SAR (Synthetic Aperture Radar) image quality enhancement method, device and equipment for target recognition

Abstract

The invention relates to a target recognition-oriented SAR image quality enhancement method, device and equipment. The method comprises the steps of generating a simulated low-quality data set by performing low-quality simulation on a high-quality data set, repairing and enhancing the simulated low-quality data set and the low-quality data set by adopting a plurality of image enhancement methods to obtain repaired and enhanced data sets, and evaluating image quality of the two types of data sets by utilizing full-reference and non-reference indexes. Training and testing a target recognition model through a sample set and a repaired and enhanced data set to obtain comprehensive accuracy; analyzing the correlation degree between the comprehensive accuracy and each quality evaluation result by adopting a correlation coefficient; screening key quality evaluation results; constructing a new loss function based on the key results; and training an image enhancement recognition model and applying the model to SAR image enhancement. According to the method, the target identification performance is improved by associating the quality index and the identification effect optimization enhancement process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of SAR target recognition, in particular to a SAR image quality enhancement method, device and equipment for target recognition. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing system based on coherent imaging principle. Unlike optical sensors, SAR's operation is not affected by light and weather conditions, with reliable all-weather and day-night observation capabilities. These advantages make SAR indispensable in disaster monitoring, ocean surveillance and other related fields. In recent years, the development of deep learning has further promoted the rapid development of SAR Automatic Target Recognition (ATR) technology. However, SAR-ATR based on deep learning faces challenges from the unique characteristics of SAR images, especially the inherent speckle noise and imaging-induced degradation represented by low contrast. Speckle noise is caused by phase difference in the coherent summation process of backscattering echoes, which manifests as a dramatic fluctuation in gray value in SAR images, thus reducing the edge and texture features of the target. At the same time, low contrast makes it difficult to fully present the structural details and blurs the key target and background boundaries. These two factors together mask the target information in SAR images, hinder feature extraction, and ultimately affect recognition accuracy. Therefore, various despeckling and contrast enhancement methods have been proposed to alleviate these problems, aiming to improve SAR image quality and thus improve downstream recognition performance.

[0003] However, existing despeckling and contrast enhancement techniques, although improving image quality indicators and visual features through plug-and-play preprocessing means, are seriously disconnected from the requirements of specific recognition, such as speckle suppression that may eliminate key texture details and excessive contrast stretching that may amplify clutter interference, resulting in poor SAR-ATR performance and even harmful effects, thus affecting the performance of downstream recognition tasks. SUMMARY

[0004] Therefore, it is necessary to provide a SAR image quality enhancement method, device and equipment for target recognition that can effectively improve the performance of downstream target recognition tasks.

[0005] A SAR image quality enhancement method for target recognition, the method comprising:

[0006] obtaining a sample set of SAR images, the sample set including a high-quality data set and a low-quality data set;

[0007] According to the high-quality data set, a simulated low-quality data set is obtained by low-quality simulation, and a plurality of image enhancement methods are used to perform image restoration and image enhancement on the simulated low-quality data set and the low-quality data set respectively, so as to obtain a repaired data set and an enhanced data set;

[0008] The quality of the images in the repaired data set and the enhanced data set is evaluated by using full-reference indicators and no-reference indicators, so as to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set;

[0009] The target recognition model in the downstream task is trained and tested by using the sample set, the repaired data set and the enhanced data set, so as to obtain the comprehensive accuracy of the target recognition model;

[0010] By using the correlation coefficient, the correlation degree between the comprehensive accuracy and each quality evaluation indicator in the full-reference indicator quality evaluation result set and the no-reference indicator quality evaluation result set is obtained, and the quality evaluation indicator with the highest correlation degree with the comprehensive accuracy is taken as the key quality evaluation indicator;

[0011] A new loss function is constructed according to the key quality evaluation indicator, the image enhancement recognition model is trained by using the new loss function, and the SAR image is enhanced by using the trained image enhancement recognition model.

[0012] In one embodiment, a speckle noise is introduced into each high-quality image in the high-quality data set to simulate low quality, so as to obtain a simulated low-quality data set.

[0013] In one embodiment, the quality of the images in the repaired data set and the enhanced data set is evaluated by using full-reference indicators and no-reference indicators, so as to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set, which includes:

[0014] The quality of the repaired data is evaluated by using the full-reference indicators and taking the high-quality data set as a reference, so as to obtain the full-reference indicator quality evaluation result set;

[0015] The quality of the repaired data and the enhanced data set is evaluated by using the no-reference indicators, so as to obtain the no-reference indicator quality evaluation result set.

[0016] In one embodiment, the full-reference indicators include: mean absolute error indicator, mean square error indicator, root mean square error indicator, signal-to-noise ratio indicator, peak signal-to-noise ratio indicator, energy signal-to-noise ratio indicator, speckle gain indicator, merit factor indicator, structural similarity index indicator and multi-scale structural similarity index indicator;

[0017] The no-reference indicators include: an equivalent visual number indicator, an image mean indicator, a ratio mean indicator, an edge protection index indicator, an edge retention degree indicator based on a mean ratio, a target clutter ratio indicator, an entropy, a contrast, a contrast enhancement index indicator, a Tenengrad gradient indicator, and an enhancement metric evaluation indicator.

[0018] In one embodiment, the target recognition model in the downstream task is trained and tested using the sample set, the repaired data set, and the enhanced data set to obtain a comprehensive accuracy of the target recognition model, including:

[0019] The SAR recognition model is trained using the high-quality data set, the low-quality data set, and the enhanced data set to obtain a trained SAR recognition model.

[0020] The SAR recognition model is tested using the simulated low-quality data set, the repaired data set, the low-quality data set, and the enhanced data set to obtain the comprehensive accuracy of the SAR recognition model.

[0021] In one embodiment, when the correlation coefficient is used to determine the correlation degree between the comprehensive accuracy and each quality evaluation indicator in the full-reference indicator quality evaluation result set and the no-reference indicator quality evaluation result set, respectively, Pearson, Spearman, and Kendall correlation coefficients are used.

[0022] In one embodiment, the new loss function is represented as:

[0023]

[0024] In the above formula, F S , N S respectively represent the full-reference key quality evaluation result and the non-reference key quality indicator, L represents the quality enhancement loss function, and λ represents an adjustable weight.

[0025] The application also provides a SAR image quality enhancement device for target recognition, which includes:

[0026] A sample data set acquisition module is configured to acquire a sample set of SAR images, wherein the sample set includes a high-quality data set and a low-quality data set.

[0027] A sample data processing module is configured to obtain a simulated low-quality data set by simulating low quality according to the high-quality data set, and to perform image restoration and image enhancement on the simulated low-quality data set and the low-quality data set using multiple image enhancement methods to obtain a repaired data set and an enhanced data set.

[0028] an image evaluation module configured to evaluate the images in the repaired data set and the enhanced data set using full-reference indicators and no-reference indicators to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set;

[0029] a model comprehensive accuracy obtaining module configured to train and test a target recognition model in a downstream task using the sample set, the repaired data set, and the enhanced data set to obtain a comprehensive accuracy of the target recognition model;

[0030] a key quality evaluation result obtaining module configured to obtain, using a correlation coefficient, a quality evaluation indicator having the highest correlation degree with the comprehensive accuracy as a key quality evaluation indicator according to the correlation degrees between the comprehensive accuracy and each quality evaluation indicator in the full-reference indicator quality evaluation result set and the no-reference indicator quality evaluation result set, respectively;

[0031] an image enhancement module training and recognition module configured to construct a new loss function according to the key quality evaluation indicator, train an image enhancement recognition model using the new loss function, and perform image enhancement on a SAR image using the trained image enhancement recognition model.

[0032] A computer device including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the SAR image quality enhancement method for target recognition when executing the computer program.

[0033] A computer readable storage medium having a computer program stored thereon, the computer program implementing the steps in the SAR image quality enhancement method for target recognition when executed by a processor.

[0034] The SAR image quality enhancement method for target recognition, the device, and the apparatus described above generate a simulated low-quality data set by simulating low-quality data from a high-quality data set, repair and enhance the simulated low-quality data set and the low-quality data set using multiple image enhancement methods to obtain repaired and enhanced data sets, evaluate the image quality of the two types of data sets using full-reference and no-reference indicators to obtain corresponding evaluation result sets, train and test a target recognition model using the sample set, the repaired data set, and the enhanced data set to obtain a comprehensive accuracy, analyze the correlation degrees between the comprehensive accuracy and each quality evaluation result using a correlation coefficient, filter key quality evaluation results, construct a new loss function based on the key results, train an image enhancement recognition model, and use the image enhancement recognition model for SAR image enhancement. This method optimizes the enhancement process by correlating quality indicators and recognition effects to improve target recognition performance. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of the SAR image quality enhancement method for target recognition in one embodiment.

[0036] Figure 2 Flowchart for constructing loss function for training image enhancement model in one embodiment;

[0037] Figure 3 Illustration of restoration results of 2S1 target in MSTAR with L=1 speckle noise under two quantization strategies and different methods in one experiment, wherein, Figure 3 (a) is the Lee method, Figure 3 (b) is the Wavelet method, Figure 3 (c) is the PPB method, Figure 3 (d) is the SAR-BM3D method, Figure 3 (e) is the HE method, Figure 3 (f) is the CLAHE method, Figure 3 (g) is the Retinex method, Figure 3 (h) is the S2V method, Figure 3 (i) is the SAR-CAM method, Figure 3 (j) is the SAR-Trans method;

[0038] Figure 4 Illustration of quality enhancement results of 59-1TC target in GMVT under two quantization strategies and different methods in one experiment, wherein, Figure 4 (a) is the Lee method, Figure 4 (b) is the Wavelet method, Figure 4 (c) is the PPB method, Figure 4 (d) is the SAR-BM3D method, Figure 4 (e) is the HE method, Figure 4 (f) is the CLAHE method, Figure 4 (g) is the Retinex method, Figure 4 (h) is the S2V method, Figure 4 (i) is the SAR-CAM method, Figure 4 (j) is the SAR-Trans method;

[0039] Figure 5 Illustration of recognition results of six baseline models under different quality enhancement methods and two quantization strategies on MSTAR SOC in one experiment, wherein, Figure 5 (a) represents the radar chart of the recognition performance of the six baseline models based on random initialization weight training under the clamping quantization condition of the quality enhancement algorithm, Figure 5(b) Radar chart of recognition performance of the six baseline models trained based on pre-trained weights under the condition of quality enhancement algorithm with clipping quantization, Figure 5 (c) Radar chart of recognition performance of the six baseline models trained based on randomly initialized weights under the condition of quality enhancement algorithm with linear quantization, Figure 5 (d) Radar chart of recognition performance of the six baseline models trained based on pre-trained weights under the condition of quality enhancement algorithm with linear quantization;

[0040] Figure 6 Fig. 6 shows a radar chart of recognition results of six baseline models under different quality enhancement methods and two quantization strategies on GMVT in an experiment, wherein, Figure 6 (a) Radar chart of recognition performance of the six baseline models trained based on randomly initialized weights under the condition of quality enhancement algorithm with clipping quantization, Figure 6 (b) Radar chart of recognition performance of the six baseline models trained based on pre-trained weights under the condition of quality enhancement algorithm with clipping quantization, Figure 6 (c) Radar chart of recognition performance of the six baseline models trained based on randomly initialized weights under the condition of quality enhancement algorithm with linear quantization, Figure 6 (d) Radar chart of recognition performance of the six baseline models trained based on pre-trained weights under the condition of quality enhancement algorithm with linear quantization;

[0041] Figure 7 Fig. 7 shows a structural block diagram of a SAR image quality enhancement device for target recognition in an embodiment

[0042] Figure 8 Fig. 8 shows an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0044] In view of the fact that the existing SAR image quality enhancement methods generally lack a systematic evaluation framework matching the requirements of the recognition task, which leads to the problem that although these image enhancement techniques improve the overall visual clarity, the influence on different downstream recognition models is still unclear, and the disconnection between image quality indicators and recognition performance leads to the problem of failing to improve the recognition accuracy of downstream tasks, in the present application, as shown in Figure 1 Fig. 1, a SAR image quality enhancement method for target recognition is provided, which specifically comprises the following steps:

[0045] Step S100, a sample set of SAR images is obtained, and the sample set includes a high-quality data set and a low-quality data set.

[0046] Step S110, a simulated low-quality data set is obtained by simulating low quality according to the high-quality data set, and image restoration and image enhancement are performed on the simulated low-quality data set and the low-quality data set respectively by using multiple image enhancement methods, to obtain a repaired data set and an enhanced data set.

[0047] Step S120, the images in the repaired data set and the enhanced data set are quality evaluated by using full-reference indicators and no-reference indicators, to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set.

[0048] Step S130, the sample set, the repaired data set and the enhanced data set are used to train and test a target recognition model in a downstream task, to obtain a comprehensive accuracy of the target recognition model.

[0049] Step S140, by using a correlation coefficient, a quality evaluation indicator with the highest correlation degree with the comprehensive accuracy is obtained as a key quality evaluation indicator, according to the correlation degrees between the comprehensive accuracy and each quality evaluation indicator in the full-reference indicator quality evaluation result set and the no-reference indicator quality evaluation result set.

[0050] Step S150, a new loss function is constructed according to the key quality evaluation indicator, the image enhancement recognition model is trained by using the new loss function, and the SAR image is enhanced by using the trained image enhancement recognition model.

[0051] In the present application, three key parts affecting image enhancement are cooperated together, i.e. SAR image quality enhancement, SAR image quality evaluation and downstream recognition task, through a multi-dimensional correlation analysis framework as shown in Figure 2 Firstly, high-quality and low-quality data sets are constructed according to visual clarity and baseline recognition performance. Then, full-reference and no-reference indicators are applied to quantify the recovery fidelity of simulated degradation and the robustness in real-world scenarios. Then, the recognition model is trained and tested on the original data set (high-quality and low-quality data set), the degraded data set (simulated low-quality data set) and the quality enhanced data set (repaired data set and enhanced data set), to establish a task-driven accuracy benchmark. The correlation coefficient is used to strictly analyze the statistical correlation between these indicators and the recognition accuracy. Therefore, the key indicators with significant statistical relationship with the recognition performance can be determined. Subsequently, the corresponding correlation constraint loss function is constructed according to the key indicators, and noise suppression and feature preservation are jointly optimized, so as to establish a connection between SAR image quality enhancement and recognition task.

[0052] In step S100, first, a SAR image dataset is acquired, in which a plurality of measured SAR images are included, and the measured SAR images are classified according to visual clarity and baseline recognition performance to obtain high-quality SAR images and low-quality SAR images, and corresponding datasets are constructed.

[0053] In step S110, image quality evaluation is performed in two scenarios of simulated noise evaluation and real noise evaluation. In the simulated noise evaluation scenario, noise in a real scenario is simulated (such as the speckle noise commonly seen in SAR images), the effects of different enhancement methods after repair are tested, and the repair is quantified by an index. In the real noise evaluation scenario, without artificially adding noise, the natural low-quality data is directly used to test the quality of the enhanced method after repair, and the effect on real poor data is verified.

[0054] In this embodiment, in the simulated noise evaluation scenario, first, speckle noise is artificially introduced and added to each high-quality image in the high-quality dataset to simulate low quality, and a simulated low-quality dataset is obtained. Then, a plurality of image enhancement methods are used to repair each simulated low-quality image in the simulated low-quality dataset, and a repaired image is obtained. Here, each simulated low-quality image corresponds to a corresponding repaired image for each image enhancement method, and a repaired dataset is constructed.

[0055] Further, in the real noise evaluation scenario, a plurality of image enhancement methods are directly used to enhance the real low-quality images in the low-quality dataset, and enhanced images are obtained. Similarly, here each low-quality image corresponds to a corresponding enhanced image for each image enhancement method, and an enhanced dataset is constructed.

[0056] In this embodiment, the image enhancement method can use a traditional enhancement method, such as speckle noise suppression, contrast enhancement, multi-view processing, or wavelet transform, a deep learning enhancement method, such as unsupervised learning, supervised learning, or attention mechanism, or a combination of traditional and deep learning methods. Specifically, according to the actual situation, a plurality of methods suitable for the current situation can be selected for the above image repair and image enhancement.

[0057] In step S120, in the two evaluation scenarios, full-reference indicators and no-reference indicators are used to evaluate the quality of the images in the repaired dataset and the enhanced dataset, and full-reference indicator quality evaluation result sets and no-reference indicator quality evaluation result sets are obtained, including: using full-reference indicators, the high-quality dataset is used as a reference to evaluate the quality of the repaired data to obtain a full-reference indicator quality evaluation result set. Using no-reference indicators, the quality of the repaired data and the enhanced dataset is directly evaluated to obtain a no-reference indicator quality evaluation result set.

[0058] In the embodiment, for the simulation noise evaluation scene, the repaired image and the original quality image are scored by using the full-reference index, i.e. quality evaluation, such as calculating the definition and the structural similarity, and the closer to the original image, the higher the score. At the same time, the repaired image is scored by using the no-reference index, for example, evaluating whether there is a blocking effect and noise residue in the image, and automatically judging the quality.

[0059] In the embodiment, for the real noise evaluation scene, since there is no reference image, the enhanced image is scored by using the no-reference index.

[0060] Further, after all the quality evaluation results are summarized, the full-reference index quality evaluation result set and the no-reference index quality evaluation result set are obtained.

[0061] In the embodiment, the full-reference index includes the mean absolute error index (MAE), the mean square error index (MSE), the root mean square error index (RMSE), the signal-to-noise ratio index (SNR), the peak signal-to-noise ratio index (PSNR), the energy signal-to-noise ratio index (ESNR), the debanding gain index (DG), the figure of merit index (FoM), the structural similarity index (SSIM), and the multi-scale structural similarity index (MS-SSIM), and the calculation process of each full-reference index is shown in Table 1.

[0062] Specifically, the indexes from MAE to FoM mainly measure the pixel-level error of the entire image by comparing each pixel, and FoM is used to evaluate the edge retention by comparing the edge map. SSIM simulates the sensitivity of the human visual system to structural information by comparing brightness, contrast and structure, so as to evaluate the perceptual similarity. MS-SSIM further extends this method by multi-scale decomposition, and integrates the structural similarity measurement at different resolutions, so as to better keep consistent with the layered visual perception.

[0063] In the embodiment, since there are challenges in acquiring pairs of high-quality and low-quality SAR images in the real world, the full-reference indicators are not applicable on real SAR images. Therefore, the no-reference indicators are adopted. The no-reference indicators include: equivalent number of looks indicator (ENL), mean of image indicator (MoI), mean of ratio indicator (MoR), edge protection index indicator (EPI), edge preservation degree based on ratio of mean indicator (EPD-ROA), target clutter ratio indicator (TCR), entropy, contrast, contrast improvement index indicator (CII), Tenengrad gradient indicator (TEN), and enhanced metric evaluation indicator (EME), and the calculation processes of the no-reference indicators are shown in Table 2. Unlike the full-reference indicators which require accurate pixel-level correspondence with the ground truth, the selected no-reference indicators evaluate the key features of SAR images, such as speckle suppression, edge preservation, sharpness, and target-background contrast. This paradigm shift enables comprehensive and reference-independent quality evaluation based on physically interpretable features.

[0064] Table 1 Full-reference indicators

[0065]

[0066] Table 2 No-reference indicators

[0067]

[0068] Although these indicators can quantitatively evaluate the enhancement effect, they often deviate from the needs of the recognition task. Traditional indicators such as ENL focus on the quality perceived by humans, but often suppress features related to recognition. This mismatch leads to a paradoxical result. Methods that obtain high scores on indicators often only bring limited accuracy improvement, while moderate enhancement, although with lower scores, can better preserve key features. These contradictions highlight the need for the relevance-driven bridge between quality evaluation and recognition performance in the method.

[0069] In step S130, the evaluation results are driven by the downstream recognition task, linking image quality indicators to recognition performance, so that task-specific accuracy benchmarks can be generated to measure the actual utility of enhancement strategies.

[0070] In the embodiment, the target recognition model in the downstream task is trained and tested using the sample set, the repaired data set, and the enhanced data set to obtain the comprehensive accuracy of the target recognition model, including: training the SAR recognition model using the high-quality data set, the low-quality data set, and the enhanced data set to obtain the trained SAR recognition model. Then, the SAR recognition model is tested using the simulated low-quality data set, the repaired data set, the low-quality data set, and the enhanced data set to obtain the comprehensive accuracy of the SAR recognition model.

[0071] Specifically, when training the target recognition model, the basic data, i.e., the high-quality data set, the enhanced data, i.e., the low-quality data set, and the enhanced data set, are mixed and then the target recognition model is trained.

[0072] Specifically, the trained target recognition model is tested in two scenarios, i.e., a simulated repair scenario and a real enhancement scenario. In the simulated repair scenario, the simulated low-quality data and the repaired data set are used to test the target recognition model. In the real enhancement scenario, the low-quality data set and the enhanced data set are used to test the target recognition model.

[0073] In step S140, in order to establish a task-driven quality indicator, the accuracy is matched with the corresponding image quality indicator for pair-wise correlation analysis, and the correlation analysis is used to screen the quality evaluation results, wherein the correlation analysis adopts Pearson, Spearman and Kendall correlation coefficients.

[0074] In this embodiment, the correlation coefficient is a statistical measure used to quantify the degree of correlation between two variables, with a value range of [-1, 1]. The larger the absolute value, the stronger the correlation, and the sign indicates the direction of the relationship. The Pearson correlation coefficient is used to evaluate the linear relationship of continuous data under the normal assumption:

[0075]

[0076] In formula (1), x and y are the values of two variables, x and y are the average values of the corresponding variables, and n is the number of variables. The Pearson correlation coefficient requires linearity and homoscedasticity, and is sensitive to outliers.

[0077] In this embodiment, the Spearman correlation coefficient uses rank-transformed data to evaluate monotonic trends:

[0078]

[0079] In formula (2), where di represents the rank difference of the variable, this non-parametric method can tolerate non-normal distribution and outliers.

[0080] In this embodiment, the Kendall correlation coefficient measures the rank consistency through the consistent (C) and inconsistent (D) pairs:

[0081]

[0082] The Kendall correlation coefficient is more suitable for small data sets or cases where there are rank ties, and it emphasizes interpretability rather than computational efficiency.

[0083] In this embodiment, Pearson assumes parametric conditions, while Spearman and Kendall prioritize robustness to non-linear or non-Gaussian data.

[0084] Further, through comprehensive correlation analysis, the full-reference key quality indicators, i.e., key full-reference indicators F S and non-reference key quality indicators, i.e., non-reference indicators N S , closely related to the recognition performance can be determined. In the optimization process, these indicators essential to recognition replace the traditional MSE and TV loss, thereby realizing joint supervision of noise suppression and feature preservation. The proposed correlation constraint loss, i.e., the new loss function, can be simply expressed as:

[0085]

[0086] In formula (4), F S , N S represent full-reference key quality evaluation results and non-reference key quality indicators, respectively, L represents a quality enhancement loss function, and λ represents an adjustable weight. Among them, F S restores the image, N S retains discriminative texture. This hybrid scheme bridges the gap between perceptual quality improvement and identification-specific requirements, achieving simultaneous improvement of image quality and recognition performance.

[0087] In this embodiment, the target recognition model described above is not limited to a certain target recognition network, and the specific target recognition model is a target recognition model adopted by a downstream task in actual application. Similarly, the image enhancement model is not limited to a certain model, and the method is applicable to any model with image enhancement capability.

[0088] In this embodiment, a source code that can achieve the above correlation constraint loss is also provided, as shown in Table 3.

[0089] Table 3

[0090]

[0091] In this paper, the effectiveness of the method is also proved by experimental data. The experimental results verify the enhancement effect under controlled simulated degradation and real low-quality scenarios, and calculate the statistical correlation between the quality indicators and the recognition accuracy.

[0092] As shown in Table 4, two data sets with different characteristics are used for system evaluation.

[0093] Table 4 Types and quantities of two data sets

[0094]

[0095] The MSTAR dataset contains 10 X-band SAR military ground target images with a resolution of 0.3 m x 0.3 m and is the most classic dataset for recognition, which can be divided into Standard Operating Conditions (SOC) and Extended Operating Conditions (EOC). Since high accuracy is easy to achieve under SOC conditions, we consider MSTAR SOC as a high-quality dataset. The network is trained on the original training set and tested on the degraded and restored test set.

[0096] The GMVT dataset is divided into a training set and a test set, with the training set having an incidence angle of 26°, 31°, 37°, and 45° (acquired in July) and the test set having an incidence angle of 15°, 31°, and 45° (acquired in March). Both sets contain 9 different target classes. The dataset is characterized by severe background clutter and fuzzy target contours, and can be considered as a low-quality dataset. To eliminate the interference related to the incidence angle, we only focus on the 31° subset in this experiment, and train and test the network on the original and quality-enhanced datasets.

[0097] Under standardized conditions, 10 SAR image quality enhancement methods were evaluated: Lee, Wavelet, PPB, SAR-BM3D, HE, CLAHE, Retinex, S2V, SAR-CAM, and SAR-Trans. Traditional methods were implemented using the default parameters in their original publications, while deep learning-based methods used their provided pre-trained weights or retrained models with the same parameter configurations as specified in their original publications. The grayscale output was normalized to the 0-255 range through two quantization strategies:

[0098] Clipped quantization:

[0099]

[0100] Linear quantization:

[0101]

[0102] Six neural networks including ResNet18, VGG16, ResNet50, EfficientNetV2, SwinTransformerV2-Tiny and ConvNext-Tiny are studied, whose weights include both random initialization and pre-training from the optical dataset ImageNet. In terms of dataset preparation, MSTAR images are centered cropped to 128x128 and resized to 256x256, while GMVT images are directly resized to 256x256. Both datasets are subjected to noise injection or quality enhancement at the 256x256 resolution stage, then uniformly center cropped to 224x224 for training and testing. The preprocessing procedure and parameter initialization of all methods are kept consistent to ensure fairness.

[0103] The spot noise with L=1 is introduced in the MSTAR SOC test set for evaluation. The evaluation results are summarized in Table 4. Under the full-reference index, SAR-BM3D performs superiorly, even exceeding the deep learning-based methods. In contrast, the contrast enhancement techniques have limited ability to restore degraded images, as can be seen from the pixel-level error rise. The no-reference index reveals the different advantages of various methods: traditional despeckling methods can better preserve fine details, while contrast enhancement algorithms can enhance the richness of texture and edge sharpness. It is worth noting that the contrast methods obtain higher ENL scores, while ENL is designed to suppress speckle, which highlights their shortcomings in the target-oriented heterogeneous SAR scene. The quantization strategy also affects image quality. Compared with linear quantization, the clip quantization can better balance noise reduction and detail preservation.

[0104] Table 4 Average full-reference and no-reference evaluation results of different enhancement methods under simulated speckle noise on the MSTAR test set under two quantization strategies

[0105]

[0106] Among them, the best and the second best results are highlighted in bold and underlined respectively.

[0107] Figure 3 The visual results under simulated speckle noise are shown. The output results of SAR-BM3D and SAR-CAM are closest to the reference image, while other despeckling methods leave residual noise or cause blurring. Contrast enhancement images increase brightness but amplify background clutter. Different quantization strategies mainly affect global brightness, with little effect on target-specific features.

[0108] Discrepancies between metrics and visual effects: PPB performs well in edge preservation but suffers from residual noise, while contrast and CII poorly correlate with perceived contrast. TEN and EME, on the other hand, align with human evaluations, reflecting fidelity in edge sharpness and contrast. Therefore, multi-metric validation is crucial to reduce inherent biases of no-reference metrics.

[0109] As shown in Table 5, evaluations on real low-quality GMVT datasets demonstrate that quantization strategies have little impact on no-reference metrics. The performance of deep learning-based methods in preserving key object details is comparable to traditional speckle reduction techniques, which is different from results on simulated noise where traditional methods exhibit higher structural fidelity.

[0110] Table 5. Average no-reference evaluation results of different quality enhancement methods on GMVT under two quantization strategies

[0111]

[0112] As Figure 4 shown, visual results highlight the salient features of different methods. Clipping quantization preserves a brightness distribution closer to the original SAR image, while linear quantization introduces artificial brightness fluctuations. All speckle reduction algorithms suffer from residual speckle noise, and PPB and SAR-Trans outputs show blurred object edges and loss of high-frequency details. S2V oversimplifies textures, producing an output that approximates binarization and blurs the boundaries between objects and clutter. Contrast enhancement methods emphasize object features but also amplify background clutter and sidelobe interference.

[0113] Discrepancies between metric evaluations and perceived quality highlight the limitations of current no-reference metrics. Although MoI, MoR, EPI, and EPD-ROA are designed to measure structural integrity, their scores fail to distinguish meaningful edges from artifacts caused by noise. For example, S2V achieves a higher EPD-ROA score by overemphasizing high-contrast edges, which mainly originate from noise rather than effective structural features. Similarly, these metrics lack sensitivity to background interference, as demonstrated by HE-enhanced images where the enhanced background edges increase the scores despite the object region being affected. These findings reinforce the necessity of shifting from general quality enhancement to a recognition-centric optimization paradigm.

[0114] As Figure 2(c) As shown, ten SAR image quality enhancement methods were applied to the degraded MSTAR SOC test set and the entire GMVT dataset. Each method generated a separate dataset. Six baseline models were selected in the experiment to represent different architectural paradigms: VGG16 represents a classic deep CNN, ResNet18 and ResNet50 represent hierarchical feature extraction based on residual learning, EfficientNetV2 represents a lightweight and scalable design, SwinV2-Tiny represents a hierarchical modeling based on Transformer, and ConvNext-Tiny represents a modernized CNN architecture inspired by Transformer. These models were trained on the original MSTAR SOC training set as well as the original and ten quality-enhanced GMVT training sets. Subsequently, these models were evaluated on simulated noisy images, real noisy images, and their quality-enhanced counterparts to assess the effectiveness of quality enhancement methods in improving recognition accuracy.

[0115] Results on MSTAR SOC: Table 6 and Figure 5 The recognition results of six baseline models on the original, degraded, and recovered MSTAR SOC test sets are summarized. On the original dataset, the randomly initialized VGG16 achieved the highest accuracy, while the ConvNext-Tiny performed poorly, which could be due to its complex architecture causing overfitting. On clean data, pre-trained models generally outperformed randomly initialized models. Although achieving higher recognition rates, all six models exhibited varying degrees of degradation under simulated speckle noise, which was caused by the distortion of target key textures, leading to domain shift and challenging feature extraction. Notably, randomly initialized models exhibited stronger robustness under simulated speckle noise. This is because pre-trained models inherently favor natural image features in their learned representations, resulting in domain mismatch with simulated speckle noise. In contrast, randomly initialized networks adaptively prioritize noise-robust mid-level features during training, bypassing the restrictive priors in pre-trained architectures. This flexibility enhances the network's adaptability to simulated noise patterns while maintaining discriminative ability.

[0116] Table 6 Precision of six baseline models under two initialization methods on MSTAR SOC with different quality enhancement methods and two quantization strategies

[0117]

[0118] Observations on the restored test set show that most of the speckle reduction methods improve the recognition rate compared to the degraded data, while the contrast enhancement methods even decrease the accuracy. The recognition rate recovery is highly consistent with the full-reference metric ranking. Under the clamp quantization condition, SAR-BM3D and SAR-CAM achieve the highest recognition accuracy, which is consistent with their highest full-reference score ranking. Under the linear quantization condition, SAR-CAM and SAR-Trans also achieve excellent recognition performance, which is consistent with their leading full-reference metric results. In contrast, the contrast enhancement methods have the lowest full-reference score and further decrease the accuracy relative to the degraded input.

[0119] The no-reference metrics have limited correlation with the recognition results. Although Wavelet and PPB can obtain higher no-reference scores and preserve some structural details in the presence of residual noise, S2V over-smoothes the key textures into nearly detail-free appearance. However, S2V surprisingly brings better recognition rate recovery, which is consistent with its better full-reference metric. Similarly, under the linear quantization condition, SAR-BM3D slightly outperforms SAR-Trans in the no-reference metric and preserves finer details with minimal noise, although its intensity is darker. SAR-Trans introduces noticeable blurring artifacts, but its higher recognition accuracy still matches its excellent full-reference metric despite the visual degradation.

[0120] In both quantization methods, the recognition recovery capability of clamp quantization is superior to linear quantization, but this difference seems to be due to its stronger consistency with the full-reference metric rather than the advantage of its own algorithm.

[0121] These findings suggest that the effectiveness of recognition recovery mainly depends on the performance of the full-reference metric, while the no-reference metric, visual noise residue, and detail blurring have little impact on the recognition results.

[0122] Results on GMVT: Table 7 and Figure 6 The recognition results of six baseline models on the GMVT dataset are shown, including the original version and the quality-enhanced version. In contrast to the results on the MSTAR SOC with simulated speckle noise, the accuracy of the pre-trained models is significantly higher than that of the randomly initialized networks in almost all enhancement methods. This difference is likely due to the pre-trained models utilizing hierarchical features learned from optical images, which can better generalize to real SAR challenges such as clutter and scatter variation. In contrast, the randomly initialized models lack this prior knowledge, making it difficult to distinguish targets from noisy backgrounds.

[0123] Table 7 Accuracy of six baseline models under two initialization methods on MSTAR SOC with different quality enhancement methods and two quantization strategies

[0124]

[0125] Notably, contrast enhancement methods improve the accuracy of randomly initialized VGG16 and EfficientNetV2, likely by amplifying low-level discriminative features such as edge sharpness. However, these methods degrade performance in the pre-trained setting. This suggests that pre-trained networks optimized for optical image statistics mistake artificially enhanced high-contrast artifacts as irrelevant distortions, thereby disrupting semantic alignment. On the other hand, randomly initialized models adaptively prioritize these amplified features during end-to-end training, but at the expense of limited generalization. S2V consistently degrades recognition accuracy in all cases. Its aggressive binarization eliminates nearly all texture, making the target and background indistinguishable, especially in cluttered scenes of GMVT. Figure 4 Visual results in confirm the strength of the near-planarity, highlighting the incompatibility between over-smoothing and recognition relying on texture features.

[0126] There is a clear disconnect between no-reference metric performance, visual clarity, and recognition accuracy. Methods such as SAR-BM3D, Wavelet, and SAR-CAM excel on no-reference metrics measuring edge sharpness and texture preservation, producing visually sharp output with minimal artifacts. However, despite these methods' high metric scores and visual performance, their improvement in recognition accuracy is minimal. In contrast, methods such as Lee, PPB, and SAR-Trans rank lower on no-reference metrics but achieve significant improvements in recognition rate. These methods all incorporate varying degrees of smoothing. Lee introduces slight blurring while preserving faint structural patterns; PPB suppresses fine details and produces patch-like artifacts; and SAR-Trans evens out textures, nearly eliminating high-frequency variations. Furthermore, the methods with the largest improvements in recognition accuracy are associated with the most aggressive smoothing: SAR-Trans and PPB achieve the highest accuracy gains, while Lee's weaker smoothing results in smaller gains. This pattern highlights a key bifurcation, where methods that excel at no-reference metrics and visual clarity fail to improve recognition rates, while those that sacrifice visual performance for controllable smoothness consistently outperform the former.

[0127] In terms of quantization strategies, the clamping and linear methods differed in terms of brightness and contrast, but their effects on identification were minimal.

[0128] All these results reinforce the limited impact of noise residue and detail blurring on the recognition results. In particular, moderate texture simplification suppresses noise-sensitive features while preserving key structural gradients, thus improving recognition accuracy. This suggests that controlled detail loss can stabilize the extracted features in low-quality SAR images, prioritizing robustness over visual clarity.

[0129] To address the severe disconnection between SAR image quality enhancement and recognition effectiveness, the statistical correlation between image quality indicators and recognition performance was further systematically analyzed. This study aims to identify key indicators for recognition and provide reliable guidance for recognition-oriented image quality enhancement algorithms. To ensure robustness, three correlation coefficients between image quality indicators and recognition accuracy were computed, where the Pearson coefficient represents linear relationship, the Spearman coefficient represents monotonic trend, and the Kendall coefficient represents rank consistency. Recognition performance is quantified by the average accuracy of the top three models in each quality enhancement method among six baseline networks to reduce bias towards specific architectures. Table 8 and Table 9 summarize the correlation computation results between image quality indicators and recognition accuracy.

[0130] Table 8 Correlation coefficients between different SAR image quality evaluation indicators and MSTAR SOC recognition accuracy under two initialization methods and two quantization strategies

[0131]

[0132] Table 9 Correlation coefficients between different SAR image quality evaluation indicators and GMVT recognition accuracy under two initialization methods and two quantization strategies

[0133]

[0134] Correlation analysis reveals key insights into the relationship between SAR image quality indicators and recognition performance. Full-reference indicators, especially DG and PSNR, exhibit strong positive correlation with recognition accuracy, reflecting their ability to quantify structural fidelity and noise suppression, which are crucial for target recognition. In contrast, SSIM and MS-SSIM, which aim to assess perceptual similarity, show weaker correlation, possibly due to their sensitivity to brightness and contrast changes, which have less relevance to target recognition. FoM, which evaluates edge preservation, shows moderate correlation, indicating that excessive edge sharpness does not directly improve recognition rates. MAE shows inconsistent correlation, reflecting its limited ability to capture specific spatial relationships for recognition.

[0135] No-reference metrics generally show poor or negative correlations with recognition results. ENL shows a moderately negative correlation, suggesting that ENL is not suitable for heterogeneous SAR target regions and that oversuppression of speckle noise can degrade texture, which is crucial for recognition. Metrics such as MoI and MoR, which measure the radiometric preservation of filtered results, also exhibit slightly negative trends when their values ​​approach theoretical ideals, as they are more suitable for homogeneous regions without targets. Edge and sharpness metrics, including EPD-ROA, EPI, and TEN, show significant negative correlations, which directly aligns with experimental observations showing that moderate smoothing, while degrading these metrics, can improve recognition due to the attenuation of high-frequency noise. Contrast-related metrics such as contrast, CII, and TCR show negligible positive or inconsistent correlations, suggesting that global contrast adjustment has a limited impact on recognition. In contrast, EME shows a strong negative correlation, suggesting that local overenhancement can impair recognition. Entropy shows a weak negative correlation, further emphasizing that image information complexity does not guarantee the usefulness of recognition features.

[0136] These results highlight the superiority of full-reference metrics for guiding recognition-oriented image quality enhancement. They also confirm that traditional no-reference metrics prioritize visual clarity over robustness for identifying specific features. The results demonstrate that moderate smoothing effectively balances noise suppression and feature preservation, emphasizing structural gradients essential for classification. These trends are further supported by consistency across datasets, with quantization strategies having negligible impact on relevant patterns. In summary, this analysis lays the foundation for optimizing image quality enhancement algorithms and improving the connection between image quality assessment and recognition performance.

[0137] Based on the correlation analysis, a new loss function is proposed that exploits the statistical relationship between the observed quality metrics and recognition performance to jointly optimize the image quality enhancement efficacy and the preservation of recognition-specific features. Specifically, the DG metric shows a strong positive correlation with recognition accuracy, while the TEN metric shows a clear negative correlation. To reconcile these opposing trends, the loss function is calculated as:

[0138]

[0139] In formula (5), the proposed loss function integrates two key parts to coordinate SAR image restoration and discriminative feature preservation. The first part uses the DG indicator to minimize the inpainted image. The pixel-level deviation between the true value X and X, thereby strengthening noise suppression, while maintaining the positive correlation with recognition accuracy. However, optimization driven by DG alone may retain artifacts. To alleviate this problem, we introduce a second regularization term derived from the TEN indicator, which decomposes the TEN indicator into horizontal and vertical Sobel operator directional gradients in horizontal and vertical directions to penalize high-frequency artifacts that are detrimental to recognition, which is negatively correlated with recognition accuracy. The balance between horizontal and vertical gradients is controlled by the coefficient λ x ∈ [0, 1], which is adaptive to the noise direction. The global regularization coefficient λ TEN controls the trade-off between DG-based restoration and TEN-driven suppression. By inversely weighting DG and TEN according to their correlation with recognition performance, the loss function can optimize visual clarity and recognition accuracy.

[0140] To verify the effectiveness of the proposed loss function, SAR-CAM and SAR-Trans are selected as representative supervised SAR image quality enhancement methods for implementation. Comprehensive evaluation is conducted using the identified recognition key indicators: full-reference indicator DG and no-reference indicator TEN. In addition, six baseline recognition models are adopted for task-driven evaluation under the clamped and linear quantization strategies.

[0141] Table 10 DG and task-driven evaluation results of SAR image quality enhancement methods based on the original loss function and the improved loss function on MSTAR SOC

[0142]

[0143] Top row: random initialization; bottom row: pre-training.

[0144] Table 11 SSIM and task-driven evaluation results of two SAR image quality enhancement methods GMVT based on the original loss function and the improved loss function

[0145]

[0146] As shown in Tables 10 and 11. The proposed correlation-constrained loss function has significant improvements in DG, TEN, and recognition accuracy. The experimental results reveal a consistent pattern that DG enhancement and TEN reduction synergistically improve recognition efficiency, verifying the statistical correlation identified in Section IV-E. This consistency confirms that optimizing key indicators can effectively improve recognition performance.

[0147] It is worth noting that individual inconsistencies were observed in the MSTAR evaluation, i.e., the improvement of DG occasionally deviated from the recognition results. These deviations stemmed from the correlation quantification method, which prioritized the average accuracy of the top three performing models rather than relying on a single specific network. Moreover, while DG maintained the strongest overall correlation, other full-reference metrics such as SSIM also maintained high correlations. The modified loss function caused DG values to increase while SSIM values decreased, which could potentially decrease the recognition rate overall. This highlights the inadequacy of single metric optimization, and thus the necessity of coordinating multi-metric supervision for robust enhancement. These findings collectively emphasize the necessity of matching quality metrics to specific recognition requirements.

[0148] Experiments demonstrate that by proposing a multi-dimensional correlation analysis framework, the problem of serious disconnection between SAR image quality improvement and target recognition performance is solved. Comprehensive evaluations on MSTAR and GMVT datasets confirm that SAR image quality metrics, especially the no-reference metrics that emphasize visual clarity, often conflict with specific recognition requirements. In addition, moderate smoothing also helps to improve recognition efficiency. Further correlation analysis shows that most full-reference metrics have strong correlation with recognition accuracy, among which DG shows the most significant positive correlation. On the contrary, no-reference metrics such as TEN, which aim to evaluate edge sharpness or preservation, show a clear negative correlation with recognition results. The correlation-constrained loss function proposed by the method balances noise suppression and feature preservation by identifying key metrics, achieving simultaneous improvement of perceptual quality and downstream recognition performance. The method makes collaborative enhancement between SAR image quality and recognition model possible by establishing a multi-dimensional correlation analysis framework. The research results provide basic guidance for the development of recognition-oriented SAR image quality enhancement algorithms, and promote the integration of remote sensing preprocessing and deep learning.

[0149] In the above SAR image quality enhancement method for target recognition, first, high-quality and low-quality datasets are constructed according to visual clarity and baseline recognition performance. Then, full-reference and no-reference metrics are applied to quantify the restoration fidelity of simulated degradation and robustness in real-world scenarios. Then the recognition model is trained and tested on the original dataset, the degraded dataset and the quality enhanced dataset to establish a task-driven accuracy benchmark. Pearson, Spearman and Kendall coefficients are used to rigorously analyze the statistical correlation between these metrics and recognition accuracy. Therefore, the key metrics with significant statistical relationship with recognition performance can be determined. Subsequently, a novel correlation-constrained loss function is proposed, which can jointly optimize noise suppression and feature preservation, thereby establishing a connection between SAR image quality enhancement and recognition tasks.

[0150] A multi-dimensional correlation analysis framework is proposed in this method, which is the first systematic study of the statistical correlation between SAR image quality indicators and recognition performance. Through three correlation coefficients, a robust theoretical basis is established for optimizing recognition-oriented SAR image quality enhancement algorithms. At the same time, a correlation-constrained loss function is proposed, which can suppress noise while preserving discriminative features, thereby achieving simultaneous improvement of image quality and recognition accuracy.

[0151] It should be understood that, although Figure 1 The steps in the flowchart of Figure 1 At least some of the steps in

[0152] In one embodiment, as shown in Figure 7 A target recognition-oriented SAR image quality enhancement device is provided, comprising: a sample data set acquisition module 200, a sample data processing module 210, an image evaluation module 220, a model comprehensive accuracy obtaining module 230, a key quality evaluation result obtaining module 240, and an image enhancement module training and recognition module 250, wherein:

[0153] The sample data set acquisition module 200 is used to acquire a sample set of SAR images, wherein the sample set includes a high-quality data set and a low-quality data set;

[0154] The sample data processing module 210 is used to obtain a simulated low-quality data set by simulating low quality according to the high-quality data set, and to perform image restoration and image enhancement on the simulated low-quality data set and the low-quality data set respectively using multiple image enhancement methods, thereby obtaining a repaired data set and an enhanced data set;

[0155] The image evaluation module 220 is used to perform quality evaluation on the images in the repaired data set and the enhanced data set using full-reference indicators and no-reference indicators, thereby obtaining a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set;

[0156] The model comprehensive accuracy obtaining module 230 is configured to train and test a target recognition model in a downstream task by using the sample set, the repaired data set and the enhanced data set, and to obtain a comprehensive accuracy of the target recognition model.

[0157] The key quality evaluation result obtaining module 240 is configured to obtain a quality evaluation result with the highest correlation degree with the comprehensive accuracy as a key quality evaluation result according to the correlation degree between the comprehensive accuracy and each quality evaluation result in the full-reference index quality evaluation result set and the no-reference index quality evaluation result set respectively by using the correlation coefficient.

[0158] The image enhancement module training and identification module 250 is configured to construct a new loss function according to the key quality evaluation result, train an image enhancement and identification model by using the new loss function, and perform image enhancement on a SAR image by using the trained image enhancement and identification model.

[0159] The specific limitations of the SAR image quality enhancement device for target recognition can refer to the limitations of the SAR image quality enhancement method for target recognition described above, and will not be described here. Each module in the SAR image quality enhancement device for target recognition described above can be realized by software, hardware and a combination thereof in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to each module by the processor.

[0160] In one embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram thereof can be as shown in FIG. 8. Figure 8 The computer device includes a processor, a memory, a network interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a SAR image quality enhancement method for target recognition. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad or mouse, etc.

[0161] Those skilled in the art can understand that, Figure 8The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0162] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0163] Obtaining a sample set of SAR images, the sample set including a high-quality data set and a low-quality data set;

[0164] Simulating low-quality data sets from the high-quality data set, and performing image restoration and image enhancement on the simulated low-quality data sets and the low-quality data sets using multiple image enhancement methods to obtain repaired data sets and enhanced data sets;

[0165] Using full-reference indicators and no-reference indicators to evaluate the quality of the images in the repaired data sets and the enhanced data sets to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set;

[0166] Training and testing a target recognition model in a downstream task using the sample set, the repaired data sets, and the enhanced data sets to obtain a comprehensive accuracy rate of the target recognition model;

[0167] Using a correlation coefficient to determine the degree of correlation between the comprehensive accuracy rate and each quality evaluation result in the full-reference indicator quality evaluation result set and the no-reference indicator quality evaluation result set, and obtaining a quality evaluation result with the highest degree of correlation with the comprehensive accuracy rate as a key quality evaluation result;

[0168] Constructing a new loss function based on the key quality evaluation result, training an image enhancement recognition model using the new loss function, and performing image enhancement on SAR images using the trained image enhancement recognition model.

[0169] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0170] Obtaining a sample set of SAR images, the sample set including a high-quality data set and a low-quality data set;

[0171] According to the high-quality data set, a simulated low-quality data set is obtained by low-quality simulation, and a plurality of image enhancement methods are used to perform image restoration and image enhancement on the simulated low-quality data set and the low-quality data set respectively, so as to obtain a repaired data set and an enhanced data set;

[0172] The repaired data set and the enhanced data set are used to perform quality evaluation on the images in the repaired data set and the enhanced data set, so as to obtain a full-reference index quality evaluation result set and a no-reference index quality evaluation result set;

[0173] The sample set, the repaired data set and the enhanced data set are used to train and test a target recognition model in a downstream task, so as to obtain a comprehensive accuracy of the target recognition model;

[0174] According to the correlation between the comprehensive accuracy and each quality evaluation result in the full-reference index quality evaluation result set and the no-reference index quality evaluation result set, a quality evaluation result with the highest correlation with the comprehensive accuracy is obtained as a key quality evaluation result by using a correlation coefficient;

[0175] A new loss function is constructed according to the key quality evaluation result, the image enhancement recognition model is trained by using the new loss function, and the trained image enhancement recognition model is used to perform image enhancement on the SAR image.

[0176] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0177] Any combination of the technical features in the above embodiments can be made. For the sake of brevity, the foregoing description has not described all possible combinations of the technical features in the above embodiments, however, as long as the combination of the technical features does not contradict, it should be considered within the scope of the present disclosure.

[0178] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. A SAR image quality enhancement method for target recognition, characterized in that: The method comprises: Acquire a sample set of SAR images, wherein the sample set includes a high-quality data set and a low-quality data set; Performing low-quality simulation on the high-quality data set to obtain a simulated low-quality data set, and performing image restoration and image enhancement on the simulated low-quality data set and the low-quality data set using a plurality of image enhancement methods to obtain a restored data set and an enhanced data set; Using full-reference indicators and no-reference indicators, perform quality assessment results on the images in the repaired dataset and the enhanced dataset to obtain a full-reference indicator quality assessment result set and a no-reference indicator quality assessment result set; Using the sample set, the repaired dataset, and the enhanced dataset to train and test the target recognition model in the downstream task, and obtain the comprehensive accuracy of the target recognition model; Using the correlation coefficient, according to the degree of correlation between the comprehensive accuracy and each quality assessment indicator in the full reference indicator quality assessment result set and the quality assessment result set without reference indicators, the quality assessment indicator with the highest correlation with the comprehensive accuracy is obtained as the key quality assessment indicator; A new loss function is constructed according to the key quality evaluation index, an image enhancement recognition model is trained using the new loss function, and the trained image enhancement recognition model is used to perform image enhancement on the SAR image.

2. The SAR image quality enhancement method for target recognition according to claim 1, characterized in that: Speckle noise is introduced into each high-quality image in the high-quality data set to simulate low quality, thereby obtaining a simulated low-quality data set.

3. The SAR image quality enhancement method for target recognition according to claim 2, characterized in that: The full-reference index and the no-reference index are used to perform quality assessment on the images in the repaired dataset and the enhanced dataset, and a full-reference index quality assessment result set and a no-reference index quality assessment result set are obtained, including: Using full reference indicators, the quality of the repaired data is evaluated with reference to the high-quality data set to obtain the full reference indicator quality evaluation result set; The quality of the repaired data and the enhanced data set is directly evaluated using a no-reference indicator to obtain the no-reference indicator quality evaluation result set.

4. The SAR image quality enhancement method for target recognition according to claim 3, characterized in that: The full reference indicators include: mean absolute error indicator, mean square error indicator, root mean square error indicator, signal-to-noise ratio indicator, peak signal-to-noise ratio indicator, energy signal-to-noise ratio indicator, despeckle gain indicator, figure of merit indicator, structural similarity index indicator and multi-scale structural similarity index indicator; The non-reference indicators include: equivalent view count indicator, image mean indicator, ratio mean indicator, edge protection index indicator, mean ratio-based edge preservation indicator, target clutter ratio indicator, entropy, contrast, contrast enhancement index indicator, Tenengrad gradient indicator and enhancement metric evaluation indicator.

5. The SAR image quality enhancement method for target recognition according to claim 4, characterized in that: The target recognition model in the downstream task is trained and tested using the sample set, the repaired dataset, and the enhanced dataset to obtain the comprehensive accuracy of the target recognition model, including: Using the high-quality data set, the low-quality data set, and the enhanced data set to train the SAR recognition model to obtain a trained SAR recognition model; The SAR recognition model is tested using the high-quality data set, the repaired data set, the low-quality data set, and the enhanced data set to obtain the comprehensive accuracy of the SAR recognition model.

6. The SAR image quality enhancement method for target recognition according to claim 5, characterized in that: When using the correlation coefficient, the Pearson, Spearman and Kendall correlation coefficients are used according to the degree of correlation between the comprehensive accuracy and the quality assessment indicators in the full reference indicator quality assessment result set and the quality assessment result set without reference indicators.

7. The SAR image quality enhancement method for target recognition according to claim 6, characterized in that: The new loss function is expressed as: In the above formula, F S 、N S They represent the full-reference key quality assessment results and the non-reference key quality indicators, L represents the quality enhancement loss function, and λ represents the adjustable weight.

8. A SAR image quality enhancement device for target recognition, characterized in that: The device comprises: A sample data set acquisition module is used to acquire a sample set of SAR images, wherein the sample set includes a high-quality data set and a low-quality data set; a sample data processing module, configured to perform low-quality simulation on the high-quality data set to obtain a simulated low-quality data set, and to perform image restoration and image enhancement on the simulated low-quality data set and the low-quality data set using a plurality of image enhancement methods, respectively, to obtain a restored data set and an enhanced data set; An image evaluation module is used to perform quality evaluation on the images in the restored dataset and the enhanced dataset using full-reference indicators and no-reference indicators, to obtain a full-reference indicator quality evaluation result set and a no-reference indicator quality evaluation result set; A model comprehensive accuracy obtaining module is used to train and test the target recognition model in the downstream task using the sample set, the repaired data set, and the enhanced data set to obtain the comprehensive accuracy of the target recognition model; a key quality assessment result obtaining module, configured to obtain, by using the correlation coefficient, the quality assessment indicator with the highest correlation with the comprehensive accuracy rate as the key quality assessment indicator, based on the correlation degree between the comprehensive accuracy rate and each quality assessment indicator in the full reference indicator quality assessment result set and the quality assessment result set without reference indicators; The image enhancement module training and recognition module is used to construct a new loss function based on key quality assessment indicators, use the new loss function to train the image enhancement recognition model, and use the trained image enhancement recognition model to enhance the SAR image.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images