RESEARCH APPROACH

The research approach adopted for the study consists of seven phases which includes as shown in Figure 1:

1. Pre-acquisition
2. data pre-processing
3. Raw images collection
4. Image pre-processing
5. Naming of images
6. Dataset Repository
7. Performance Evaluation

The different phases in the study are discussed in the sections below.

PRE-ACQUISITION

The volunteers are given brief orientation on how their data will be managed and used for research purposes only. After the volunteers agrees, a consent form is given to be read and signed. The sample of the consent form filled by the volunteers is shown in Figure 1.

The capturing of images was started with the setup of the imaging device. The camera is set up on a tripod stand in stationary position at the height 90 from the floor and distance 20cm from the subject.

EAR IMAGE ACQUISITION

Image acquisition is an action of retrieving image from an external source for further processing. The image acquisition is purely a hardware dependent process by capturing unprocessed images of the volunteers using a professional camera. This was acquired through a subject posing in front of the camera. It is also a process through which digital representation of a scene can be obtained. This representation is known as an image and its elements are called pixels (picture elements). The imaging sensor/camera used in this study is a Canon E0S 60D professional camera which is placed at a distance of 3 feet form the subject and 20m from the ground.

This is the first step in this project to achieve the project’s aim of developing an occlusion and pose sensitive image dataset for black ear recognition. (OPIB ear dataset). To achieve the objectives of this study, a set of black ear images were collected mostly from undergraduate students at a public University in Nigeria.

The image dataset required is captured in two scenarios:

1. uncontrolled environment with a surveillance camera

The image dataset captured is purely black ear with partial occlusion in a constrained and unconstrained environment.

2. controlled environment with professional cameras

The ear images captured were from black subjects in controlled environment. To make the OPIB dataset pose invariant, the volunteers stand on a marked positions on the floor indicating the angles at which the imaging sensor was captured the volunteers’ ear. The capturing of the images in this category requires that the subject stand and rotates in the following angles 60^o, 30^o and 0^o towards their right side to capture the left ear and then towards the left to capture the right ear (Fernando et al., 2017) as shown in Figure 4. Six (6) images were captured per subject at angles 60^o, 30^o and 0^o for the left and right ears of 152 volunteers making a total of 907 images (five volunteers had 5 images instead of 6, hence folders 34, 22, 51, 99 and 102 contain 5 images).

To make the OPIB dataset occlusion and pose sensitive, partial occlusion of the subject’s ears were simulated using rings, hearing aid, scarf, earphone/ear pods, etc. before the images are captured.

Figure 1: Sample of Subject’s Consent Form for the OPIB ear dataset

RAW IMAGE COLLECTION

The ear images were captured using a digital camera which was set to JPEG because if the camera format is set to raw, no processing will be applied, hence the stored file will contain more tonal and colour data. However, if set to JPEG, the image data will be processed, compressed and stored in the appropriate folders.

IMAGE PRE-PROCESSING

The aim of pre-processing is to improve the quality of the images with regards to contrast, brightness and other metrics. It also includes operations such as: cropping, resizing, rescaling, etc. which are important aspect of image analysis aimed at dimensionality reduction. The images are downloaded on a laptop for processing using MATLAB.

Image Cropping

The first step in image pre-processing is image cropping. Some irrelevant parts of the image can be removed, and the image Region of Interest (ROI) is focused. This tool provides a user with the size information of the cropped image. MATLAB function for image cropping realizes this operation interactively by waiting for a user to specify the crop rectangle with the mouse and operate on the current axes. The output images of the cropping process are of the same class as the input image.

Naming of OPIB Ear Images

The OPIB ear images were labelled based on the naming convention formulated from this study as shown in Figure 5. The images are given unique names that specifies the subject, the side of the ear (left or right) and the angle of capture. The first and second letters (SU) in the image names is block letter simply representing subject for subject 1-to-n in the dataset, while the left and right ears is distinguished using L1, L2, L3 and R1, R2, R3 for angles 60⁰, 30⁰ and 0⁰_, respectively as shown in Table 1.

Table 1: Naming Convention for OPIB ear images

OPIB EAR DATASET EVALUATION

The prominent challenges with the current evaluation practices in the field of ear biometrics are the use of different databases, different evaluation matrices, different classifiers that mask the feature extraction performance and the time spent developing framework (Abaza et al., 2013; Emeršič et al., 2017).

The toolbox provides environment in which the evaluation of methods for person recognition based on ear biometric data is simplified. It executes all the dataset reads and classification based on ear descriptors.

DESCRIPTION OF OPIB EAR DATASET

OPIB ear dataset was organised into a structure with each folder containing 6 images of the same person. The images were captured with both left and right ear at angle 0, 30 and 60 degrees. The images were occluded with earing, scarves and headphone etc. The collection of the dataset was done both indoor and outdoor. The dataset was gathered through the student at a public university in Nigeria. The percentage of female (40.35%) while Male (59.65%). The ear dataset was captured through a profession camera Nikon D 350. It was set-up with a camera stand where an individual captured in a process order. A total number of 907 images was gathered.

The challenges encountered in term of gathering students for capturing, processing of the images and annotations. The volunteers were given a brief orientation on what their ear could be used for before, it was captured, for processing. It was a great task in arranging the ear (dataset) into folders and naming accordingly.

Table 2: Overview of the OPIB Ear Dataset

Matlab

## Overview

Matlab is a dataset for object detection tasks - it contains Face annotations for 220 images.

## Getting Started

You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.

  ## License

  This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).

==Description Pictures of objects belonging to 101 categories. About 40 to 800 images per category. Most categories have about 50 images. Collected in September 2003 by Fei-Fei Li, Marco Andreetto, and Marc Aurelio Ranzato. The size of each image is roughly 300 x 200 pixels. We have carefully clicked outlines of each object in these pictures, these are included under the Annotations.tar . There is also a matlab script to view the annotaitons, show_annotations.m . How to use the dataset If you are using the Caltech 101 dataset for testing your recognition algorithm you should try and make your results comparable to the results of others. We suggest training and testing on fixed number of pictures and repeating the experiment with different random selections of pictures in order to obtain error bars. Popular number of training images: 1, 3, 5, 10, 15, 20, 30. Popular numbers of testing images: 20, 30. See also the discussion below. When you report your results please keep track of wh

This data release provides remotely sensed data, field measurements, and MATLAB code associated with an effort to produce image-derived velocity maps for a reach of the Sacramento River in California's Central Valley. Data collection occurred from September 16-19, 2024, and involved cooperators from the Intelligent Robotics Group from the National Aeronautics and Space Administration (NASA) Ames Research Center and the National Oceanographic and Atmospheric Administration (NOAA) Southwest Fisheries Science Center. The remotely sensed data were obtained from an Uncrewed Aircraft System (UAS) and are stored in Robot Operating System (ROS) .bag files. Within these files, the various data types are organized into ROS topics including: images from a thermal camera, measurements of the distance from the UAS down to the water surface made with a laser range finder, and position and orientation data recorded by a Global Navigation Satellite System (GNSS) receiver and Inertial Measurement Unit (IMU) during the UAS flights. This instrument suite is part of an experimental payload called the River Observing System (RiOS) designed for measuring streamflow and further detail is provided in the metadata file associated with this data release. For the September 2024 test flights, the RiOS payload was deployed from a DJI Matrice M600 Pro hexacopter hovering approximately 270 m above the river. At this altitude, the thermal images have a pixel size of approximately 0.38 m but are not geo-referenced. Two types of ROS .bag files are provided in separate zip folders. The first, Baguettes.zip, contains "baguettes" that include 15-second subsets of data with a reduced sampling rate for the GNSS and IMU. The second, FullBags.zip, contains the full set of ROS topics recorded by RiOS but have been subset to include only the time ranges during which the UAS was hovering in place over one of 11 cross sections along the reach. The start times are included in the .bag file names as portable operating system interface (posix) time stamps. To view the data within ROS .bag files, the Foxglove Studio program linked below is freely available and provides a convenient interface. Note that to view the thermal images, the contrast will need to be adjusted to minimum and maximum values around 12,000 to 15,000, though some further refinement of these values might be necessary to enhance the display. To enable geo-referencing of the thermal images in a post-processing mode, another M600 hexacopter equipped with a standard visible camera was deployed along the river to acquire images from which an orthophoto was produced: 20240916_SacramentoRiver_Ortho_5cm.tif. This orthophoto has a spatial resolution of 0.05 m and is in the Universal Transverse Mercator (UTM) coordinate system, Zone 10. To assess the accuracy of the orthophoto, 21 circular aluminum ground control targets visible in both thermal and RGB (red, green, blue) images were placed in the field and their locations surveyed with a Real-Time Kinematic (RTK) GNSS receiver. The coordinates of these control points are provided in the file SacGCPs20240916.csv. Please see the metadata for additional information on the camera, the orthophoto production process, and the RTK GNSS survey. The thermal images were used as input to Particle Image Velocimetry (PIV) algorithms to infer surface flow velocities throughout the reach. To assess the accuracy of the resulting image-derived velocity estimates, field measurements of flow velocity were obtained using a SonTek M9 acoustic Doppler current profiler (ADCP). These data were acquired along a series of 11 cross sections oriented perpendicular to the primary downstream flow direction and spaced approximately 150 m apart. At each cross section, the boat from which the ADCP was deployed made four passes across the channel and the resulting data was then aggregated into mean cross sections using the Velocity Mapping Toolbox (VMT) referenced below (Parsons et al., 2013). The VMT output was further processed as described in the metadata and ultimately led to a single comma delimited text file, SacAdcp20240918.csv, with cross section numbers, spatial coordinates (UTM Zone 10N), cross-stream distances, velocity vector components, and water depths. To assess the sensitivity of thermal image velocimetry to environmental conditions, air and water temperatures were recorded using a pair of Onset HOBO U20 pressure transducer data loggers set to record pressure and temperature. Deploying one data logger in the air and one in the water also provided information on variations in water level during the test flights. The resulting temperature and water level time series are provided in the file HoboDataSummary.csv with a one-minute sampling interval. These data sets were used to develop and test a new framework for mapping flow velocities in river channels in approximately real time using images from an UAS as they are acquired. Prototype code for implementing this approach was developed in MATLAB and is also included in the data release as a zip folder called VelocityMappingCode.zip. Further information on the individual functions (*.m files) included within this folder is available in the metadata file associated with this data release. The code is provided as is and is intended for research purposes only. Users are advised to thoroughly read the metadata file associated with this data release to understand the appropriate use and limitations of the data and code provided herein.

The data consist of a file system-based data of Raj 3765 variety of wheat. There are twenty-four chlorophyll fluorescence images (Control and Drought) that have been captured for a period of sixty days. A total of (1440 x 2) images are used for this research work.

The generated chlorophyll fluorescence set is subjected to various pre-processing operations including noise removal, contrast enhancement. Pre-processed images are then segmented using novel segmentation approach “C-fit Kmeans” segmentation to increase water stress detection accuracy for automation procedure. 23 GLCM Texture features are identified from the dataset listed following:

1. Autocorrelation: (out.autoc)
2. Contrast: matlab (out.contr)
3. Correlation: matlab (out.corrm) 4.Correlation: (out.corrp) 5.Cluster Prominence: (out.cprom)
4. Cluster Shade: (out.cshad) 7.Dissimilarity: (out.dissi)
5. Energy: matlab (out.energ)
6. Entropy: (out.entro)
7. Homogeneity: matlab (out.homom)
8. Homogeneity: (out.homop)
9. Maximum probability: (out.maxpr)
10. Sum of sqaures: Variance (out.sosvh)
11. Sum average (out.savgh)
12. Sum variance (out.svarh)
13. Sum entropy (out.senth)
14. Difference variance (out.dvarh)
15. Difference entropy (out.denth)
16. Information measure of correlation1 (out.inf1h)
17. Informaiton measure of correlation2 (out.inf2h)
18. Inverse difference (INV) is homom (out.homom)
19. Inverse difference normalized (INN) (out.indnc)
20. Inverse difference moment normalized (out.idmnc)

These variables then undergo various statistical processes viz. correlation, factor, and Clutsring analyzing to identify the key drought detection variables suited best for water stress which in-turn help to build (AIM) Agronomical Inferencing Model for the wheat crop to understand the behavioral change in texture variables in the presence of stress.

The dataset has been produced using MATLAB GLCM libraries https://in.mathworks.com/help/images/ref/graycomatrix.html

Motivation

The goal of introducing the Rescaled CIFAR-10 dataset is to provide a dataset that contains scale variations (up to a factor of 4), to evaluate the ability of networks to generalise to scales not present in the training data.

The Rescaled CIFAR-10 dataset was introduced in the paper:

[1] A. Perzanowski and T. Lindeberg (2025) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, Journal of Mathematical Imaging and Vision, 67(29), https://doi.org/10.1007/s10851-025-01245-x.

with a pre-print available at arXiv:

[2] Perzanowski and Lindeberg (2024) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, arXiv preprint arXiv:2409.11140.

Importantly, the Rescaled CIFAR-10 dataset contains substantially more natural textures and patterns than the MNIST Large Scale dataset, introduced in:

[3] Y. Jansson and T. Lindeberg (2022) "Scale-invariant scale-channel networks: Deep networks that generalise to previously unseen scales", Journal of Mathematical Imaging and Vision, 64(5): 506-536, https://doi.org/10.1007/s10851-022-01082-2

and is therefore significantly more challenging.

Access and rights

The Rescaled CIFAR-10 dataset is provided on the condition that you provide proper citation for the original CIFAR-10 dataset:

[4] Krizhevsky, A. and Hinton, G. (2009). Learning multiple layers of features from tiny images. Tech. rep., University of Toronto.

and also for this new rescaled version, using the reference [1] above.

The data set is made available on request. If you would be interested in trying out this data set, please make a request in the system below, and we will grant you access as soon as possible.

The dataset

The Rescaled CIFAR-10 dataset is generated by rescaling 32×32 RGB images of animals and vehicles from the original CIFAR-10 dataset [4]. The scale variations are up to a factor of 4. In order to have all test images have the same resolution, mirror extension is used to extend the images to size 64x64. The imresize() function in Matlab was used for the rescaling, with default anti-aliasing turned on, and bicubic interpolation overshoot removed by clipping to the [0, 255] range. The details of how the dataset was created can be found in [1].

There are 10 distinct classes in the dataset: “airplane”, “automobile”, “bird”, “cat”, “deer”, “dog”, “frog”, “horse”, “ship” and “truck”. In the dataset, these are represented by integer labels in the range [0, 9].

The dataset is split into 40 000 training samples, 10 000 validation samples and 10 000 testing samples. The training dataset is generated using the initial 40 000 samples from the original CIFAR-10 training set. The validation dataset, on the other hand, is formed from the final 10 000 image batch of that same training set. For testing, all test datasets are built from the 10 000 images contained in the original CIFAR-10 test set.

The h5 files containing the dataset

The training dataset file (~5.9 GB) for scale 1, which also contains the corresponding validation and test data for the same scale, is:

cifar10_with_scale_variations_tr40000_vl10000_te10000_outsize64-64_scte1p000_scte1p000.h5

Additionally, for the Rescaled CIFAR-10 dataset, there are 9 datasets (~1 GB each) for testing scale generalisation at scales not present in the training set. Each of these datasets is rescaled using a different image scaling factor, 2^k/4, with k being integers in the range [-4, 4]:

cifar10_with_scale_variations_te10000_outsize64-64_scte0p500.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte0p595.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte0p707.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte0p841.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte1p000.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte1p189.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte1p414.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte1p682.h5
cifar10_with_scale_variations_te10000_outsize64-64_scte2p000.h5

These dataset files were used for the experiments presented in Figures 9, 10, 15, 16, 20 and 24 in [1].

Instructions for loading the data set

The datasets are saved in HDF5 format, with the partitions in the respective h5 files named as
('/x_train', '/x_val', '/x_test', '/y_train', '/y_test', '/y_val'); which ones exist depends on which data split is used.

The training dataset can be loaded in Python as:

with h5py.File(`

x_train = np.array( f["/x_train"], dtype=np.float32)
x_val = np.array( f["/x_val"], dtype=np.float32)
x_test = np.array( f["/x_test"], dtype=np.float32)
y_train = np.array( f["/y_train"], dtype=np.int32)
y_val = np.array( f["/y_val"], dtype=np.int32)
y_test = np.array( f["/y_test"], dtype=np.int32)

We also need to permute the data, since Pytorch uses the format [num_samples, channels, width, height], while the data is saved as [num_samples, width, height, channels]:

x_train = np.transpose(x_train, (0, 3, 1, 2))
x_val = np.transpose(x_val, (0, 3, 1, 2))
x_test = np.transpose(x_test, (0, 3, 1, 2))

The test datasets can be loaded in Python as:

with h5py.File(`

x_test = np.array( f["/x_test"], dtype=np.float32)
y_test = np.array( f["/y_test"], dtype=np.int32)

The test datasets can be loaded in Matlab as:

x_test = h5read(`

The images are stored as [num_samples, x_dim, y_dim, channels] in HDF5 files. The pixel intensity values are not normalised, and are in a [0, 255] range.

Motivation

The goal of introducing the Rescaled Fashion-MNIST dataset is to provide a dataset that contains scale variations (up to a factor of 4), to evaluate the ability of networks to generalise to scales not present in the training data.

The Rescaled Fashion-MNIST dataset was introduced in the paper:

[1] A. Perzanowski and T. Lindeberg (2025) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, Journal of Mathematical Imaging and Vision, 67(29), https://doi.org/10.1007/s10851-025-01245-x.

with a pre-print available at arXiv:

[2] Perzanowski and Lindeberg (2024) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, arXiv preprint arXiv:2409.11140.

Importantly, the Rescaled Fashion-MNIST dataset is more challenging than the MNIST Large Scale dataset, introduced in:

[3] Y. Jansson and T. Lindeberg (2022) "Scale-invariant scale-channel networks: Deep networks that generalise to previously unseen scales", Journal of Mathematical Imaging and Vision, 64(5): 506-536, https://doi.org/10.1007/s10851-022-01082-2.

Access and rights

The Rescaled Fashion-MNIST dataset is provided on the condition that you provide proper citation for the original Fashion-MNIST dataset:

[4] Xiao, H., Rasul, K., and Vollgraf, R. (2017) “Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms”, arXiv preprint arXiv:1708.07747

and also for this new rescaled version, using the reference [1] above.

The data set is made available on request. If you would be interested in trying out this data set, please make a request in the system below, and we will grant you access as soon as possible.

The dataset

The Rescaled FashionMNIST dataset is generated by rescaling 28×28 gray-scale images of clothes from the original FashionMNIST dataset [4]. The scale variations are up to a factor of 4, and the images are embedded within black images of size 72x72, with the object in the frame always centred. The imresize() function in Matlab was used for the rescaling, with default anti-aliasing turned on, and bicubic interpolation overshoot removed by clipping to the [0, 255] range. The details of how the dataset was created can be found in [1].

There are 10 different classes in the dataset: “T-shirt/top”, “trouser”, “pullover”, “dress”, “coat”, “sandal”, “shirt”, “sneaker”, “bag” and “ankle boot”. In the dataset, these are represented by integer labels in the range [0, 9].

The dataset is split into 50 000 training samples, 10 000 validation samples and 10 000 testing samples. The training dataset is generated using the initial 50 000 samples from the original Fashion-MNIST training set. The validation dataset, on the other hand, is formed from the final 10 000 images of that same training set. For testing, all test datasets are built from the 10 000 images contained in the original Fashion-MNIST test set.

The h5 files containing the dataset

The training dataset file (~2.9 GB) for scale 1, which also contains the corresponding validation and test data for the same scale, is:

fashionmnist_with_scale_variations_tr50000_vl10000_te10000_outsize72-72_scte1p000_scte1p000.h5

Additionally, for the Rescaled FashionMNIST dataset, there are 9 datasets (~415 MB each) for testing scale generalisation at scales not present in the training set. Each of these datasets is rescaled using a different image scaling factor, 2^k/4, with k being integers in the range [-4, 4]:

fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p500.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p595.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p707.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p841.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p000.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p189.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p414.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p682.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte2p000.h5

These dataset files were used for the experiments presented in Figures 6, 7, 14, 16, 19 and 23 in [1].

Instructions for loading the data set

The datasets are saved in HDF5 format, with the partitions in the respective h5 files named as
('/x_train', '/x_val', '/x_test', '/y_train', '/y_test', '/y_val'); which ones exist depends on which data split is used.

The training dataset can be loaded in Python as:

with h5py.File(`

x_train = np.array( f["/x_train"], dtype=np.float32)
x_val = np.array( f["/x_val"], dtype=np.float32)
x_test = np.array( f["/x_test"], dtype=np.float32)
y_train = np.array( f["/y_train"], dtype=np.int32)
y_val = np.array( f["/y_val"], dtype=np.int32)
y_test = np.array( f["/y_test"], dtype=np.int32)

We also need to permute the data, since Pytorch uses the format [num_samples, channels, width, height], while the data is saved as [num_samples, width, height, channels]:

x_train = np.transpose(x_train, (0, 3, 1, 2))
x_val = np.transpose(x_val, (0, 3, 1, 2))
x_test = np.transpose(x_test, (0, 3, 1, 2))

The test datasets can be loaded in Python as:

with h5py.File(`

x_test = np.array( f["/x_test"], dtype=np.float32)
y_test = np.array( f["/y_test"], dtype=np.int32)

The test datasets can be loaded in Matlab as:

x_test = h5read(`

The images are stored as [num_samples, x_dim, y_dim, channels] in HDF5 files. The pixel intensity values are not normalised, and are in a [0, 255] range.

There is also a closely related Fashion-MNIST with translations dataset, which in addition to scaling variations also comprises spatial translations of the objects.

Most outdoor vision systems can be influenced by rainy weather conditions. We present a single-image rain removal method, called ResDerainNet.The proposed network can automatically detect rain streaks and remove them. Based on the deep convolutional neural networks (CNN), we learn the mapping relationship between rainy and residual images from data. Furthermore, for training, we synthesize rainy images considering various rain models. Specifically, we mainly focus on the composite models as well as orientations and scales of rain streaks. In summary, we make following contributions; - A residual deep network is introduced to remove rain noise. Unlike the plane deep network which learns the mapping relationship between noisy and clean images, we learn the relationship between rainy and residual images from data. This speeds up the training process and improves the de-raining performance.

• An automatic rain noise generator is introduced to obtain synthetic rain noise. Most de-raining methods create rain noise by using Photoshop. Since synthetic rain noise has many parameters, it is difficult to automatically adjust these parameters. In our method, we can easily change some parameters on MATLAB, which saves time and effort to get natural rain noise.

• A combination of linear additive composite model and screen blend model is proposed to make synthetic rainy images. In order for the training network to be applicable to a wide range of rainy images, only one composite model is not enough. Our experimental results show that a combination of these models achieves better performance than using either model.

This dataset is compose by the images of the paper Online automatic anomaly detection for photovoltaic systems using thermography imaging and low rank matrix decomposition and the images from a Github repository of technodivesh. The authors of the first dataset shared it with 120 images as a supplemental material along with some Matlab codes. The author of the second dataset shared his data on Github due to a personal project he carried out. As the original datasets only had thermal (IR) images, I've annotate each one of them, enabling the use of instance/semantic segmentation to detect PV modules/arrays and classify it's condition (defected or not).

1. Context

Both large and small photovoltaic systems are susceptible to failures in their equipment, especially in modules due to operational stresses that are exposed and errors during the installation process of these devices. Although numerous internal and external factors originate these failures, the common phenomenon presented by several of them is hot spots on module defective area. The immediate impact is perceptible in the reduction of the generated power and, in the long term, in the reduction of the useful life of the equipment due to the high temperatures presented. The preventive maintenance method for recognizing this phenomenon is the use of thermography images in inspections of photovoltaic modules. Through this procedure, faulty modules are immediately identified with failures at an early stage due to their high heat signatures compared to the others, captured by cameras with infrared sensors. Currently, the use of this type of camera attached to drones stands out for providing an increase in the inspection area and a reduction in its execution time.

To understand more about this, read these reports by International energy agency (IEA): - Review of failures of PV modules; - Review of IR and EL images applications for PV systems.

1.1 Thermal Images

In thermography area, FLIR is the company that stands out the most, with it's equipments being the most used. It's influence in this field has ensured that his proprietary image formats are the most studied, through the use of reverse engineering (such as those published by the Exiftool website and some electronics forums). Most of your cameras currently produce images in R-JPG (radiometric JPG) format, where radiometric data is embedded in their metadata. If there is a common digital camera attached to the thermal camera, its image will also be stored in this metadata.

1.2 Image annotation

For image annotation, I used the Flyr Python library to extract the radiometric data from each image metadata (in R-JPG format) and convert it into an 8-bit image. This is important as the actual thermal image (embedded in the metadata of the image pulled from the camera) is always smaller than the original one (image pulled from the camera). Also, the original image always has some object that obstructs a relevant part of the scene, like the FLIR logo.

After this, I used Grid annotation tool 2 to demarcate the photovoltaic modules. This tool was created by Lukas Bommes to precisely help annotate objects very close to each other, in a grid structure as showed by the image bellow.

https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F5048762%2F60d6fb64038475ca48cd7df5c475ac70%2Fgrid-annotation-tool-example.png?generation=1670540579132887&alt=media" alt="grid-annotation-tool-example">

The annotation is in JSON format and if you look into this tool repository, you'll notice that I've changed some keys. I removed the "image" key, changed "grid cells" to "instances", removed "id" and "truncated" to "defected modules" in order to highlight which modules are defective. Therefore, each annotation has the bellow structure:

{
 "instances": [
  {
   "corners": [
    {"x": 181.98136043148975, "y": 180.87500292991302},
    {"x": 80.99062551749896, "y": 176.94569670353974},
    {"x": 78.42819076221744, "y": 294.24826550087346},
    {"x": 183.0432587489509, "y": 294.49812289825013}
   ],
   "center": {"x": 131.1108588650393, "y": 236.64177200814407},
   "defected_modules": true
 ...

This is a preclinical positron emission tomography (PET) dataset containing the list-mode data of a NEMA image quality phantom measured with the preclinical Siemens Inveon PET scanner. Included are the list-mode datafile (.lst), sinogram file (.scn) created by the Siemens Inveon Acquisition workplace (IAW) software, MAP-OSEM3D reconstruction (.img) created by IAW, scatter correction sinogram (_sct.scn) created by IAW and the attenuation correction UMAP-file (.img) created by IAW. All the corresponding header files are included that contain all the relevant information, with the exception of reading the binary list-mode data. For documentation on reading the list-mode binary data, please ask Siemens.

No normalization data is included in this dataset. You can, however, use the normalization data from Preclinical PET data.

This dataset can be used in the OMEGA software, including the list-mode data, to import the data to MATLAB/Octave, create sinograms from the list-mode data and reconstruct the imported data. For help on using the dataset with OMEGA, see the wiki.

The CT data, that was used to create the UMAP-file, is available from https://zenodo.org/record/4646835.

The measurement data was collected by Jarmo Teuho.

This repository contains raw data and all scripts for the in situ sequencing processing pipeline (Mats Nilsson Lab, Stockholm University) and downstream analysis applied in the manuscript “Spatial and temporal localization of immune transcripts defines hallmarks and diversity in the tuberculosis granuloma”. Raw images:All 16bit tiff images are for 12 weeks_section1Format is baseX_cY_ORGX indicates the hybridization round (1-4)Y indicates the channel: 1: DAPI2: FITC (T)3: Cy3 (G)4: TexR (C)5: Cy5 (A)6: AF750 (Anchor primer)

Raw data in folders:

Lung csv files

Bacteria csv files

DAPI for plotting

HE folder

Scripts folder:

Matlab scripts

Cellprofiler pipelines

Identification and plotting of transcripts:

For all matlab scripts: Download the “Matlab scripts” folder, add lib to MATLAB path. Except MATLAB, no additional Mathworks product is required. Tested on R2017b.

InSituSequencing.m is the top-level script processing sequencing images to positional visualization of decoded transcripts. Use raw images for 12weeks_section1 as input images (others are available on request) . After Tiling in "InsituSequencing.m", process tiled images in the cell profiler pipeline “Blob identification”. Run decode and threshold in "InSituSequencing.m" to generate csv files containing position and intensity of each identified signal.csv files for all lung scans (3 per time point) are in “lung csv files” folder and can be plotted on DAPI images (10% of original size) found in “DAPI for plotting” folder using Plotting global in "InSituSequencing.m". High resolution H&E scans of in situ-sequenced lungs for lung section per time point are in the “HE folder” at 50% of original size. For all images 1 pixel corresponds to 0.325 mm.

Identification and plotting of transcripts in given proximity to bacteria:

Use the cellprofiler pipeline “Bacteria identification” instead of “Blob identification” to identify signal in indicated distances from identified bacteria. The folder “bacteria csv files” contains identified signals in the indicated distances to identified bacteria. Input images are available on request.

Downstream analysis (Matlab Scripts folder)

DensityEstimation.m was used to display not absolute reads but a kernel density estimation of a certain gene in a 2log scale.

ROI_draw_onImage.m was applied to extract reads from annotated regions. Pictures of annotations can be found in the manuscript supplementary figure S1.

HexbinClustering.m performed an unsupervised clustering (kmeans) of spatial data with a given number of distinct clusters in a given radius.Table 1-3 contain sequences of used specific primers, padlock probes and detection oligos.

This folder contains a subset of hyperspectral images from the "HyperPistachio: Hyperspectral images of pistachios with varying aflatoxin levels" dataset.

The folder includes three subfolders for 30 hyperspectral images of pistachios with three different level of aflatoxin contamination Less Than 8ug/kg (low), Greater Than 160 ug/kg (medium), and Greater Than 300 ug/kg (high)., each 10 images. Both the image file in band interleaved format (.bil) and the header (.hdr) file have been provided for each image.

3. An example MATLAB file for reading and displaying the hyperspectral images named 'Read_HS_Image'. Make sure to download and install the Hyperspectral Image Processing Toolbox on MATLAB first before running codes.

TechnicalRemarks: Overview of the folder structure: These folders contain the actual optical images we took and that are the basis of our image cross-correlation (nomenclature is as in the paper): Sample1 - Copper Sample2 - Gold grains Sample3 - Array 1 Sample4 - Array 1 coated Sample5 - Array 2 coated Computer generated data each of these folders contains two subfolders 'drift' and 'step' containing the data referring to the drift measurement and the step-function measurement depicted in Fig.5 of the paper. The folder 'Comuter generated data' additionally contains a folder 'supplementary figureS1 data', which contains the data for said figure. The specific parameters are mentioned in the folder name (see numbers behind 'sig_x' and 'amplitude') These folders contain the matlab scripts needed to perform the image cross-correlation and create the computer generated data. Matlab scripts Digital Image Correlation These folders contain .stl files of the printed marker arrays

This repository contains the code and data underlying the publication "Computational 3D resolution enhancement for optical coherence tomography with a narrowband visible light source" in Biomedical Optics Express 14, 3532-3554 (2023) (doi.org/10.1364/BOE.487345).

The reader is free to use the scripts and data in this depository, as long as the manuscript is correctly cited in their work. For further questions, please contact the corresponding author.

Description of the code and datasets

Table 1 describes all the Matlab and Python scripts in this depository. Table 2 describes the datasets. The input datasets are the phase corrected datasets, as the raw data is large in size and phase correction using a coverslip as reference is rather straightforward. Processed datasets are also added to the repository to allow for running only a limited number of scripts, or to obtain for example the aberration corrected data without the need to use python. Note that the simulation input data (input_simulations_pointscatters_SLDshape_98zf_noise75.mat) is generated with random noise, so if this is overwritten de results may slightly vary. Also the aberration correction is done with random apertures, so the processed aberration corrected data (exp_pointscat_image_MIAA_ISAM_CAO.mat and exp_leaf_image_MIAA_ISAM_CAO.mat) will also slightly change if the aberration correction script is run anew. The current processed datasets are used as basis for the figures in the publication. For details on the implementation we refer to the publication.

Table 1: The Matlab and Python scripts with their description


    Script name
    Description


    MIAA_ISAM_processing.m
    This scripts performs the DFT, RFIAA and MIAA processing of the phase-corrected data that can be loaded from the datasets. Afterwards it also applies ISAM on the DFT and MIAA data and plots the results in a figure (via the scripts plot_figure3, plot_figure5 and plot_simulationdatafigure).


    resolution_analysis_figure4.m
    This figure loads the data from the point scatterers (absolute amplitude data), seeks the point scatterrers and fits them to obtain the resolution data. Finally it plots figure 4 of the publication.


    fiaa_oct_c1.m, oct_iaa_c1.m, rec_fiaa_oct_c1.m, rfiaa_oct_c1.m 
    These four functions are used to apply fast IAA and MIAA. See script MIAA_ISAM_processing.m for their usage.


    viridis.m, morgenstemning.m
    These scripts define the colormaps for the figures.


    plot_figure3.m, plot_figure5.m, plot_simulationdatafigure.m
    These scripts are used to plot the figures 3 and 5 and a figure with simulation data. These scripts are executed at the end of script MIAA_ISAM_processing.m.


    Python script: computational_adaptive_optics_script.py
    Python script that applied computational adaptive optics to obtain the data for figure 6 of the manuscript.


    Python script: zernike_functions2.py
    Python script that gives the values and carthesian derrivatives of the Zernike polynomials.


    figure6_ComputationalAdaptiveOptics.m
    Script that loads the CAO data that was saved in Python, analyzes the resolution, and plots figure 6.


    Python script: OCTsimulations_3D_script2.py
    Python script simulates OCT data, adds noise and saves it as .mat file for use in the matlab script above.


    Python script: OCTsimulations2.py
    Module that contains a python class that can be used to simulate 3D OCT datasets based on a Gaussian beam.


    Matlab toolbox DIPimage 2.9.zip
    Dipimage is used in the scripts. The toolbox can be downloaded online or this zip can be used.






The datasets in this Zenodo repository


    Name
    Description


    input_leafdisc_phasecorrected.mat
    Phase corrected input image of the leaf disc (used in figure 5).


    input_TiO2gelatin_004_phasecorrected.mat
    Phase corrected input image of the TiO2 in gelatin sample.


    input_simulations_pointscatters_SLDshape_98zf_noise75
    Input simulation data that, once processed, is used in figure 4.

exp_pointscat_image_DFT.mat

exp_pointscat_image_DFT_ISAM.mat

exp_pointscat_image_RFIAA.mat

exp_pointscat_image_MIAA_ISAM.mat

exp_pointscat_image_MIAA_ISAM_CAO.mat

    Processed experimental amplitude data for the TiO2 point scattering sample with respectively DFT, DFT+ISAM, RFIAA, MIAA+ISAM and MIAA+ISAM+CAO. These datasets are used for fitting in figure 4 (except for CAO), and MIAA_ISAM and MIAA_ISAM_CAO are used for figure 6.

simu_pointscat_image_DFT.mat

simu_pointscat_image_RFIAA.mat

simu_pointscat_image_DFT_ISAM.mat

simu_pointscat_image_MIAA_ISAM.mat

    Processed amplitude data from the simulation dataset, which is used in the script for figure 4 for the resolution analysis.

exp_leaf_image_MIAA_ISAM.mat

exp_leaf_image_MIAA_ISAM_CAO.mat

    Processed amplitude data from the leaf sample, with and without aberration correction which is used to produce figure 6.

exp_leaf_zernike_coefficients_CAO_normal_wmaf.mat

exp_pointscat_zernike_coefficients_CAO_normal_wmaf.mat

    Estimated Zernike coefficients and the weighted moving average of them that is used for the computational aberration correction. Some of this data is plotted in Figure 6 of the manuscript.


    input_zernike_modes.mat
    The reference Zernike modes corresponding to the data that is loaded to give the modes the proper name.

exp_pointscat_MIAA_ISAM_complex.mat

exp_leaf_MIAA_ISAM_complex

    Complex MIAA+ISAM processed data that is used as input for the computational aberration correction.

The dataset contains two datasets, which record 200/600 images. 'GunCash80X80NO200.mat' contains one type of gun and cash '3gun80X80No600.mat' contains three types of gun and cash The data format is matlab mat file, which can be loaded and displayed through the m file "load_GunCash_plot.m". All the images are recorded as 80x80 images, but each image is recorded as a 6400x1 vector. After load the file into matlab, you will find the file "alldata", which is 6402xN and the sample number N equals 200 or 600. Note that the last two rows of "alldata" are the labels of the images. The 6401 row is the label(1/-1) to indicate the existence of a gun, The 6402 row is the label(1/-1) to indicate the existence of a cash note. We use a webcam to record images, which may depict a gun and/or cash (dollar notes). The presence or absence of a gun is the public hypothesis, and the presence or absence of cash is the private hypothesis. The gun and the cash are randomly translated and rotated in each 80x80 grayscale image. The target is to detect the public hypothesis and make the private hypothesis cannot be detected.

Analyzed DTS datasets from active heat injection experiments in Guelph, ON Canada is included. A .pdf file of images including borehole temperature distributions, temperature difference distributions, temperature profiles, and flow interpretations is included as the primary analyzed dataset. Analyzed data used to create the .pdf images are included as a matlab data file that contains the following 5 types of data: 1) Borehole Temperature (matrix of temperature data collected in the borehole), 2) Borehole Temperature Difference (matrix of temperature difference above ambient for each test), 3) Borehole Time (time in both min and sec since the start of a DTS test), 4) Borehole Depth (channel depth locations for the DTS measurements), 5) Temperature Profiles (ambient, active, active off early time, active off late time, and injection).

Different data-sets needed by the Matlab package locFISH. locFISH allows the simulation and analysis of realistic single molecule FISH (smFISH) images.

data_simulation.zip Contains all necessary data to simulated smFISH images. Specifically, the zip archive contains a library of 3D cell shapes, realistic imaging background, and a simulated PSF (Point Spread Function).

GAPDH.zip Contains the smFISH data of GAPDH and the corresponding analysis results, which were used to create the library of cell shapes provided in data_simulation.zip

For more details on these data and how do to use them, please consult the detailed user-manual provided with locFISH, available at

https://bitbucket.org/muellerflorian/fish_quant

These are the MATLAB scripts (with a ReadMe.txt) that will allow the user to point to a data file (.I16), and set up the processing script to analyze it and generate image stacks.

This dataset contains the data used to produce the paper: 'Simultaneous maximum a posteriori longitudinal PET image reconstruction' by Ellis and Reader, Physics in Medicine and Biology (2017). DOI: http://dx.doi.org/10.1088/1361-6560/aa7b49. Please see the article for a full description of methodology used to obtain this data.

The dataset comprises a number of MATLAB data files (.mat), MATLAB scripts (.m), and plain text files (.txt), corresponding to each figure in the article. Running the .m script in MATLAB for each figure will reproduce that figure approximately as it appears in the article. Furthermore, the .txt files describe the contents of the .mat data files in order to allow independent exploration of the data. Note that the function plotSparseMarker is required to be able to run fig5.m.

This work was funded by the King's College London & Imperial College London EPSRC Centre for Doctoral Training in Medical Imaging (grant number EP/L015226/1) and supported by the EPSRC grant number EP/M020142/1. This data has been made available in accordance with the EPSRC's policy framework on research data.

Lumbar Spinal Stenosis causes low back pain through pressures exerted on the spinal nerves. This can be verified by measuring the anteroposterior diameter and foraminal widths of the patient’s lumbar spine. Our goal is to develop a novel strategy for assessing the extent of Lumbar Spinal Stenosis by automatically calculating these distances from the patient’s lumbar spine MRI. Our method starts with a semantic segmentation of T1- and T2-weighted composite axial MRI images using SegNet that partitions the image into six regions of interest. They consist of three main regions-of-interest, namely the Intervertebral Disc, Posterior Element, and Thecal Sac, and three auxiliary regions-of-interest that includes the Area between Anterior and Posterior elements. A novel contour evolution algorithm is then applied to improve the accuracy of the segmentation results along important region boundaries. Nine anatomical landmarks on the image are located by delineating the region boundaries found in the segmented image before the anteroposterior diameter and foraminal widths can be measured. The performance of the proposed algorithm was evaluated through a set of experiments on the Lumbar Spine MRI dataset containing MRI studies of 515 patients. These experiments compare the performance of our contour evolution algorithm with the Geodesic Active Contour and Chan-Vese methods over 22 different setups. We found that our method works best when our contour evolution algorithm is applied to improve the accuracy of both the label images used to train the SegNet model and the automatically segmented image. The average error of the calculated right and left foraminal distances relative to their expert-measured distances are 0.28 mm (p = 0.92) and 0.29 mm (p = 0.97), respectively. The average error of the calculated anteroposterior diameter relative to their expert-measured diameter is 0.90 mm (p = 0.92). The method also achieves 96.7% agreement with an expert opinion on determining the severity of the Intervertebral Disc herniations.

This data consists of the MATLAB source code, improved label images, and composite images used in the experiment. Additional datasets and source code are necessary: - The Lumbar Spine MRI Dataset http://dx.doi.org/10.17632/k57fr854j2.2 - The Radiologists Notes for Lumbar Spine MRI Dataset http://dx.doi.org/10.17632/s6bgczr8s2.2 - The MATLAB source code for developing Ground Truth Dataset, Semantic Segmentation, and Evaluation for the Lumbar Spine MRI Dataset http://dx.doi.org/10.17632/8cp2cp7km8.2 - The original (unmodified) Label Image Data for Lumbar Spine MRI Dataset http://dx.doi.org/10.17632/zbf6b4pttk.2