The object of research is the process of suppressing multiplicative speckle noise in synthetic aperture radar (SAR) images, which significantly complicates their analysis. The problem addressed is the lack of end-to-end hybrid methods capable of spatial adaptation by integrating a mathematical model of local statistics (the Frost filter) directly into the neural network computation graph. This research is aimed at automating the process of adaptive SAR image filtering by developing the hybrid VGG16-FrostNet method. These research tasks were addressed by formulating a differentiable mathematical model of the classical Frost filter for integration into a neural network, developing an architecture based on a pretrained VGG16 (Visual Geometry Group) backbone (blocks 1–2), and integrating the Convolutional Block Attention Module (CBAM), which predicts a spatially varying damping coefficient map Amap within 0.5–10.0 for each pixel. The developed hybrid architecture includes a residual branch for detail recovery and was optimized end-to-end using a comprehensive loss function combining L1, Edge Loss (Sobel), SSIM, and attention regularization. The model was trained on synthetic data with gamma-distributed speckle (equivalent looks between 3.0 and 6.0) under typical SAR conditions. On the test set, experimental evaluation yielded a mean PSNR of 34.18 dB and SSIM of 0.97. The gain relative to the noisy image constituted 9.45 dB, and 3.36 dB in PSNR compared to the classical Frost filter with an optimal static coefficient. Edge indicators EPI = 0.8903 and FOM = 0.8340 substantiate reliable preservation of structural boundaries. It was established that the developed hybrid method provides spatially adaptive damping with interpretable attention maps, enabling its deployment in automated SAR data processing pipelines.

Speckle noise inherent to synthetic aperture radar (SAR) imagery degrades image quality and complicates automated analysis in Earth observation applications. Quantitative assessment of despeckling results requires computing quality metrics against reference images, which are unavailable in operational SAR scenarios. This paper presents a method for a priori prediction of filtered Sentinel-1 SAR image quality metrics before applying speckle noise filters. Unlike existing approaches predicting relative quality improvement, the proposed method predicts absolute values of five metrics (PSNR, WSNR, SSIM, MS-SSIM, FSIM) for a specific filter, enabling direct comparison and rational filter selection. The methodology employs transfer learning of DenseNet-121 convolutional neural network, pre-trained on ImageNet, adapted for single-channel SAR inputs through architectural modifications including input layer transformation, pooling optimization, and regression head replacement. A novel synthetic data generation pipeline utilizes histogram matching of Sentinel-2 optical images with Sentinel-1 SAR references to create training samples preserving ground truth. Dynamic gamma-distributed speckle noise addition with variable ENL ∈ [2, 6] enhances data variability and model robustness. Experiments with six classical filters (Gamma MAP, Lee, Enhanced Lee, Frost, SRAD, Kuan) demonstrate high prediction accuracy across all filter-metric combinations. The coefficient of determination R² reaches 0.997 for best combinations and exceeds 0.97 for most of the 30 trained models. Mean absolute prediction errors remain below 0.29 dB for PSNR and 0.014 for SSIM across all tested configurations. The approach enables a priori quality prediction without reference images, allowing optimization of SAR processing workflows and resource planning before resource-intensive despeckling. Contributions of Authors: Conceptualization, Raed A. and Oleksii R.; methodology, Raed A.; formal analysis, Raed A.; investigation, Raed A. and Oleksii R.; data curation, Raed A. and Oleksii R.; writing—original draft preparation, Raed A.; writing—review and editing, Raed A. and Oleksii R.; visualization, Raed A. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Data Availability Statement: The processed dataset is available in the Zenodo repository (https://zenodo.org/uploads/17253925). Python implementation code is available in the GitHub repository (https://github.com/rsenaikh/Predicting_Quality_after_Noise_Removal). Acknowledgments: The authors would like to express their sincere gratitude to the Copernicus Data Space Ecosystem for providing open access to Sentinel-1 and Sentinel-2 data. We are also grateful to reviewers and editors for their valuable comments, recommendations, and attention to the work. Conflicts of Interest: The authors declare no conflict of interest.

Digital images are increasingly used to analyze different types of objects and their localization and classification. There are many areas for using this information; it is often employed for surveillance systems, automatic driving of vehicles, or exploration of new territories. At the same time, there are a fairly large number of neural networks that allow implementation of this functionality by training them using data sets of various types and classifications. Often, data sets created with the help of unmanned aerial vehicles are frequently used for research tasks. Such datasets allow the recognition of various types of objects without direct access to them, which allows safe exploration of different territories. The use of unmanned aerial vehicles is quite common nowadays, especially in the fields of photography and videography. Many photographers use unmanned aerial vehicles to take pictures of landscapes and use automatic tracking systems for movement. Automatic movement systems and object search systems are quite sensitive to the size of the object and the quality of the search algorithm. Because of the wide applicability of this task, as well as the small amount of initial data, the topic of our work is the study of the dependence of the accuracy of localization and classification of objects on their area in images obtained using unmanned aerial vehicles. The main subject of this study is the quality of neural networks that allow obtaining information about objects, as well as research by obtaining statistical data and a test set of data on the dependence of detection accuracy on the size of the object. The goal of this study was to obtain statistics on the accuracy of localization and classification depending on the size of the object and to determine the accuracy thresholds using the obtained statistics. The task of this study is to train common neural networks with an open architecture on a set of data obtained using unmanned aerial vehicles and to determine their characteristics, particularly the dependence of recognition accuracy on the size of the object. The expected result of the work is the threshold values of the size of the object, which are permissible for a sufficiently accurate classification and localization of objects, as well as the metrics of the quality of the work of the studied neural networks. Because of this work, conclusions are given that reflect the threshold values of object sizes, on which the recognition accuracy depends.

The localization and classification of objects of different types in images is an important and actively researched topic because the designed methods and tools are exploited in a wide variety of fields, including remote sensing, security systems, and medical diagnostics. Imaging systems installed on-board unmanned aerial vehicles (UAVs) and drones have become popular recently, and they are potentially beneficial for numerous applications like mine detection, traffic control, and crowd control. Images acquired by such systems may suffer from low quality because of the use of rather cheap cameras and the necessity to transfer obtained data via communication lines with limited bandwidth, employing lossy compression. These factors can influence the quality and accuracy of object localization, which is typically negatively performed by trained neural networks. However, the intensity of the noise and distortions that can be considered acceptable, i.e. such that they do not lead to radical reduction of the performance characteristics are unclear. Given this, it is reasonable to investigate the impact of these effects on the quality of object localization and classification using a reliable data size and various noise/distortion intensities. Therefore, the research subject of this paper is the performance of object localization and classification methods for color images acquired by UAV-installed sensors. The primary focus is on the dependence of localization and classification metrics on the noise intensity, where the simulated noise mimics not only noise but also distortions due to lossy compression by modern coders. The aim of this work is to obtain adequate statistics and analyze them to build dependencies of the metrics on the intensity of distortions. The objective is to obtain conditions for which the effects of noise and distortions can be considered negligible or acceptable in practice. The second objective is to analyze the sensitivity of several modern neural network models to noise/distortions. The result is a statistical assessment of the dependence of model performance on input data quality. The conclusions are based on the statistics characterizing the model performance for the noise/distortion intensity interval. The conclusions allow the selection of the best (most robust) neural networks and the establishment of appropriate performance conditions.

The use of unmanned and manned aerial vehicles for remote object localization and classification is very common. These methods are used in various systems, ranging from territory surveys to law enforcement. Methods of object localization and classification using neural networks require a detailed study and research of the quality of their work on data that has certain specifics, such as vehicle detection. The use of neural networks to detect certain types of objects using images obtained from aircraft can also help in the study of hard-to-reach locations. Therefore, the main subject of this paper is the localization and classification of objects in images obtained using digital cameras mounted on aircraft. The main focus is on determining the accuracy of object localization and detection using selected types of neural networks, which are the most important indicators of neural network efficiency. The speed of a neural network is also an equally important characteristic as it directly affects its ability to be used in tasks that require fast object localization, such as video surveillance or automated car control systems. The main goal of this study is to study the accuracy of object localization and classification in images obtained with the help of cameras mounted on aircraft, as well as to study the speed of neural networks and determine the effectiveness of their application in real-world conditions. The objectives of this study are to train YOLO v5, SSD, and Faster RCNNs on the VisDrone dataset and to further study them on the vehicle localization dataset. The main goal of this work is to obtain statistics on the performance of neural networks trained on the VisDrone dataset. On the basis of the obtained statistics, conclusions are drawn about the effectiveness of the considered neural networks. The conclusions are drawn by considering the speed of the model, localization (IoU), and classification (Precision, Recall) metrics. Possible directions for further development of the topic under study are presented as conclusions.

The use of unmanned aerial vehicles is gaining wide popularity in various areas of research and information acquisition. More and more often, unmanned aerial vehicles are used to obtain various types of images of the Earth’s surface for its study. In particular, such data can be used in law enforcement, localization of crowds, etc. Typically, such systems operate independently of humans and provide information about objects in an automatic mode, with humans working only under the control of the aircraft. One of the main components of such systems is a neural network for localization and classification of objects, the parameters of which determine both the accuracy of the system as a whole and the design of the aircraft for shooting. In particular, the accuracy of the neural network determines the profitability of such a system, because if the accuracy is insufficient, the use of such systems will not make sense. Therefore, the main subject of research in this paper is a neural network for object localization, in particular YOLO v5, and its accuracy parameters on images obtained from unmanned aerial vehicles. The main focus of this paper is on the parameters of the neural network and the study of its metrics, which are important parameters of a trained neural network. Another important parameter for the further use of a neural network is its training parameters, as well as the data used for training. This study also pays attention to the details of the training process. The main goal of this study is to train a neural network on a selected dataset and to study the accuracy metrics of the trained neural network. The main goal of this study is to determine the dependence of localization accuracy on the area of the object, which will allow for more detailed development of unmanned systems with automatic object detection, as well as to assess the profitability of using such systems in task planning. On the basis of the data obtained, conclusions were drawn about the dependence of localization accuracy on the area of an object in images from unmanned aerial vehicles. These data can serve as a reference for unmanned aerial vehicle developers, particularly when selecting photo modules or planning the system architecture.

This chapter contains the results obtained during execution of Ukrainian-Polish Project in 2020 intended on design of methods and means for processing grayscale and multichannel images and video using visual quality metrics. The combined metrics have been

This chapter contains the results obtained during execution of Ukrainian-French Project within “Dnipro” framework intended on design of methods and means for processing multichannel remote sensing data using the trained neural networks. These means can be

This chapter contains the results obtained during execution of Ukrainian-French Project within “Dnipro” framework in 2022 intended on design of methods and means for processing multichannel remote sensing data using the trained neural networks. Neural net

Remote sensing data are widely used in numerous applications. A conventional task solved using remote sensing images is their classification. The classification maps are commonly produced by some pre-trained classifiers applied either to uncompressed or compressed images where lossy compression is often needed and employed in practice due to the necessity to reduce data volume at stages of image transfer and storage. Then, the classification accuracy depends on the characteristics of an image, a classifier, and a coder used. The main subject of this paper is the factors that determine classification accuracy. One of them is compressed image quality. We fix the quality of compressed image quality characterized by the peak signal-to-noise ratio for several coders and rely on the same training approach. Our goal is twofold. First, we would like to consider classification accuracy for two approaches to classifier training: based on undistorted data and images with simulated distortions. Second, our desire is to compare the performance of different techniques of image compression. The task of this paper is to obtain an idea is it worth training the neural network classifier for uncompressed images or images of similar quality to the quality of compressed data to be classified. The coder’s influence on classification results is of special interest as well. The main results are the following. First, the classification accuracy is almost the same for classifiers trained for uncompressed and simulated compressed data for the general distortion model. Second, there is a certain difference in the classification results for different compression techniques studied. Lightly better classification results are observed for data produced by more sophisticated (modern) coders. Experiments have been carried out for two real-life three-channel Sentinel-2 images of Kharkiv and the Kharkiv region having different complexity. Four typical classes have been considered. As a conclusion, it is possible to state that either the general model of distortions must be modified or the classifier training should be performed for data produced by the corresponding compression technique.

This study proposes VGG16-FrostNet, a hybrid method for Synthetic Aperture Radar (SAR) despeckling, and compares it against the classical Frost filter equipped with an image-specific optimal damping parameter maximizing the HaarPSI metric. A pre-trained V

Image denoising (filtering) is one of the most important procedures in image pre-processing chain. The aim of applying denoising is to reduce the noise while preserving image details and features. However, in practice, there are quite many cases when image denoising can lead to degradation of image quality which affects further thematic (high-level) image processing steps. Thus, the benefit of applying denoising is not always observed. In these cases, it is better to skip image denoising. This paper discusses decision-making on image denoising expedience based on denoising efficiency prediction for several modern filters. Analysis of visual quality for denoised images using subjective evaluation is carried out. Performance evaluation of existing image quality measures in the context of their application to assess the quality of denoised images is conducted. Finally, an algorithm for decision-making on image denoising expedience is proposed. It is shown that reliable decision on image denoising expedience can be undertaken in advance without carrying out denoising itself.

Image processing and compression algorithms might introduce various types of distortions. Thus, a question of estimating the effect of image compression on the accuracy of data classification using criteria specific only for image classification is relevant. In this paper, the effect of lossy compression on three-channel image classification using a multilayer neural net-work are analysed. It is shown that overall probability of correct classification remains almost the same for a wide range of distortions introduced by lossy compression using atomic wavelets al-though some variations of correct classification probabilities take place for particular classes depending on peculiarities of feature distributions. Experiments are carried out for multispectral Sen-tinel images of different complexity.

Multichannel systems of remote sensing provide a huge amount of data useful for different applications. However, such images occupy a large space that poses problems of processing, storage, transmission, and management. Lossy compression is widely used to decrease the size of data. In lossy compression, one has to provide a reasonable trade-off between compression ratio (CR) and introduced losses or quality of compressed data. Quality can be characterized in various ways including traditional criteria as peak signal-to-noise ratio (PSNR) or some others as well as criteria that describe efficiency of solving the final tasks of remote sensing as, e.g., probability of correct classification. In this paper, we concentrate on classification of three-channel images that can be either color images or three components of multi- or hyperspectral data acquired, e.g., by Sentinel-2 sensor. In lossy compression of color images, downscaling of color components is often applied to increase CR without essential loss of quality. The goal of this paper is to study the influence of such downscaling on classification accuracy for three-channel remote sensing data. The compression method based on atomic functions is considered since this method allows easy control of compressed image quality and its providing. The neural networks trained for distorted-free images are applied for image classification. Analysis is carried out for four images of different complexity. Based on it, practical recommendations are given.

Remote sensing images have found numerous applications nowadays. A traditional outcome or intermediate result of their processing is a classification map. Such maps are usually obtained from a pre-trained classifier and it is desired to have the produced classification maps as accurately as possible. The basic subject of this article is the factors determining this accuracy. The main among them are the quality of remote sensing data and classifier type, parameters and training approach. Image quality can be degraded due to several factors. One of them is distortions introduced by lossy compression that is widely used due to a huge volume of acquired data and the necessity to sufficiently decrease their size at transmission, storage and/or dissemination stages. Because of this, the main goal of this paper is to consider classification and lossy compression jointly. In particular, this means that the classifier learning can be performed for original (uncompressed, compressed in a lossless manner) images (if they are available) as well as for compressed data at hand (offered to a user for classification and further analysis). The task of this paper is to consider and compare these two options. The first one is the classifier learning for original images and further application to compressed data, where images can be compressed with different compression ratios while producing compressed data of different quality. The second option is the use of the classifier learning for compressed images, where compression parameters for training data can be approximately the same as for the images to which the classifier is applied. The main result is that the latter methodology can provide certain benefits compared to the classifier learning for original data if one has to classify compressed remote sensing data. Simulation data are obtained for a classifier based on a convolutional neural network. As images for training and verification, four real-life three-channel (visible range) Sentinel-2 remote sensing images of Kharkiv and Kharkiv region are employed that possess different complexity of the content and have four main classes. The practical recommendations are given. In conclusion, we can state that it is worth having classifiers trained for several degrees of compression and it is reasonable to compress complex structure images with special care.

Synthetic aperture radar (SAR) images have found numerous applications. However, further analysis of SAR images including interpretation, classification, segmentation, etc. is an extremely challenging task due to the presence of highly intensive speckle noise. Therefore, image despeckling is one of the main stages in preliminary SAR data processing. Over the past decades, a large number of different image despeckling techniques have been proposed ranging from local statistics filters to deep learning based ones. In this study, we analyze one of the most known and widely used local statistics Frost filter. Despeckling efficiency of the Frost filter significantly depends on the sliding window size and tuning (also called damping) factor. Here, we present a method for optimal parameters selection of the Frost filter for a given image based on despeckling efficiency prediction. Despeckling efficiency prediction for the Frost filter is carried out using a set of statistical and spectral input parameters and multilayer neural network. It is shown that such a prediction can be performed before applying image despeckling with a high accuracy and it is faster than despeckling itself. Both simulated speckled images and real-life Sentinel-1 SAR images have been used for extensive evaluation of the proposed method.

Radar imaging has many advantages. Meanwhile, SAR images suffer from a noise-like phenomenon called speckle. Many despeckling methods have been proposed to date but there is still no common opinion as to what the best filter is and/or what are its parameters (window or block size, thresholds, etc.). The local statistic Lee filter is one of the most popular and best-known despeckling techniques in radar image processing. Using this filter and Sentinel-1 images as a case study, we show how filter parameters, namely scanning window size, can be selected for a given image based on filter efficiency prediction. Such a prediction can be carried out using a set of input parameters that can be easily and quickly calculated and employing a trained neural network that allows determining one or several criteria of filtering efficiency with high accuracy. The statistical analysis of the obtained results is carried out. This characterizes improvements due to the adaptive selection of the filter window size, both potential and based on prediction. We also analyzed what happens if, due to prediction errors, erroneous decisions are undertaken. Examples for simulated and real-life images are presented.

A similarity search in images has become a typical operation in many applications. A presence of noise in images greatly affects the correctness of detection of similar image blocks, resulting in a reduction of efficiency of image processing methods, e.g., non-local denoising. In this paper, we study noise immunity of various distance measures (similarity metrics). Taking into account a wide variety of information content in real life images and variations of noise type and intensity. We propose a set of test data and obtain preliminary results for several typical cases of image and noise properties. The recommendations for metrics' and threshold selection are given. Fast implementation of the proposed benchmark is realized using CUDA technology.

Image denoising is a classical preprocessing stage used to enhance images. However, it is well known that there are many practical cases where different image denoising methods produce images with inappropriate visual quality, which makes an application of image denoising useless. Because of this, it is desirable to detect such cases in advance and decide how expedient is image denoising (filtering). This problem for the case of wellknown BM3D denoiser is analyzed in this paper. We propose an algorithm of decision-making on image denoising expedience for images corrupted by additive white Gaussian noise (AWGN). An algorithm of prediction of subjective image visual quality scores for denoised images using a trained artificial neural network is proposed as well. It is shown that this prediction is fast and accurate.

Synthetic aperture radars (SARs) provide a lot of images that can be used for numerous applications. A problem with acquired images is that they are corrupted by speckle which is a noise-like phenomenon with multiplicative nature. In addition, speckle is non-Gaussian and it is often spatially correlated. A typical task in SAR image processing is despeckling and many methods have been already proposed. However, most of them do not take noise spatial correlation into account during denoising. In this paper, we show how this can be done in despeckling based on discrete cosine transform. The use of frequency-dependent thresholds leads to sufficient improvement of denoising efficiency in terms of visual quality metrics. Moreover, we consider quite complex structure texture images for which noise removal is usually problematic and can lead to information loss. Comparison to the well-known local statistic Lee and Frost filters, extended DCT-based filter is carried out for different remote sensing systems including Sentinel-1 and Sentinel-2.