Low-illuminance space target contrast feature enhancement method based on self-supervised learning

Abstract

The invention discloses a low-illuminance space target comparison feature enhancement method based on self-supervised learning. The method comprises the steps of collecting a space image; carrying out preprocessing operation on the acquired space image; performing feature extraction on the data after data enhancement; performing noise suppression on the image after feature extraction; performing feature enhancement on the key region of the image after noise suppression; similar feature classification and target identification and analysis are carried out on the image after feature enhancement; and carrying out image enhancement on the image subjected to the target analysis. According to the invention, image feature enhancement driven by self-supervised learning is realized, and more comprehensive and accurate scene information is obtained; through a scale invariant feature transformation algorithm, the stability and consistency of image features under a low illumination condition are ensured, and the accuracy of feature enhancement is improved. Noise interference in a low-illumination environment is effectively suppressed, the significance of target features is enhanced, and the application range of a low-illumination image feature enhancement technology based on self-supervised learning is expanded.

Description

Technical Field

[0001] The present invention relates to a method for enhancing contrast features of spatial targets in low illumination based on self-supervised learning, and belongs to the technical field of image processing. Background Technology

[0002] In low-light environments, the detection and recognition of space targets is an important research direction in the field of image processing and computer vision. Due to insufficient light, low-light images are usually accompanied by problems such as high noise levels, low contrast, and loss of details. These factors seriously affect the accurate detection and recognition of space targets.

[0003] Traditional low-light image enhancement methods mainly rely on image enhancement techniques, such as histogram equalization, Retinex theory and its variants. These methods try to restore the details and structure of the image by adjusting the brightness and contrast of the image. However, traditional enhancement techniques have the following major problems when processing low-light images: Insufficient noise suppression: Under low-light conditions, the noise level in the image increases significantly, and traditional methods usually use fixed filters or noise models for denoising. However, this method lacks adaptability and is difficult to effectively suppress noise under different noise types and intensities, resulting in loss of target details or residual noise; Limited contrast adjustment: Although methods such as histogram equalization can improve image contrast, excessive contrast enhancement under low-light conditions may cause artifacts and edge distortion in the image, affecting the accurate recognition of spatial targets; Insufficient feature extraction: Traditional methods mainly focus on the brightness and contrast adjustment of the image, lacking effective extraction and optimization of deep image features. This results in the lack of significant features of spatial targets in complex backgrounds and low signal-to-noise ratio environments, making it difficult to achieve high-precision detection and recognition; Reliance on manual parameter adjustment: Many traditional enhancement methods require manual setting of parameters, such as the size of the filter, the intensity of contrast adjustment, etc. This not only increases the complexity of use, but also limits the generalization ability of the method in different scenarios.

[0004] In addition, in practical applications, such as military reconnaissance, night monitoring, and traffic monitoring under low light, real-time processing and high-quality enhancement of low-light images place higher demands on system performance. Traditional methods often have difficulty balancing processing efficiency and image quality in these high-demand scenarios, and cannot meet the requirements of real-time performance and high resolution.

[0005] To address the above issues, there is an urgent need for a method that can effectively enhance image contrast features under low illumination conditions, which can not only adaptively suppress noise and optimize image features, but also automatically adapt to different environmental conditions without manual parameter adjustment, thereby improving the detection accuracy and robustness of space targets. SUMMARY OF THE INVENTION

[0006] The purpose of the present invention is to provide a method for enhancing the contrast features of low-light spatial targets based on self-supervised learning. Through the self-supervised contrast learning mechanism and adaptive noise suppression technology, efficient enhancement of low-light images and accurate detection of spatial targets are achieved, overcoming the shortcomings of traditional methods in noise suppression, contrast adjustment and feature extraction, and meeting the performance requirements in high-demand application scenarios.

[0007] In order to achieve the above object, the present invention provides a method for enhancing the contrast features of low-illuminance space targets based on self-supervised learning, comprising the following steps: Step 1, collect spatial images in low-light spatial scenes; Step 2: Perform data input and data enhancement operations on the low-light spatial image collected in Step 1, including rotation, cropping, etc. Step 3, use the self-supervised contrastive learning mechanism to extract features from the image after data enhancement in Step 2; Step 4, use adaptive noise suppression mechanism to dynamically suppress the image after feature extraction in Step 3; Step 5, use the attention mechanism to enhance the features of the key areas of the image after dynamic noise suppression in Step 4; Step 6: Use density-based clustering algorithm to classify similar features of the image enhanced in Step 5 and perform target recognition and analysis; Step7: Use the non-regularized self-supervised learning mechanism to enhance the image after the target analysis in Step6.

[0008] Furthermore, the specific steps of Step 2 are: Step 2.1, perform geometric transformations including rotation, translation, cropping, scaling and flipping on the dim space image; Step 2.2, perform color space transformation including brightness adjustment, contrast adjustment, color disturbance and grayscale on the dark space image after geometric transformation, and use these enhanced samples as "positive samples", and compare the samples from different images as "negative samples".

[0009] Furthermore, the specific steps of Step 3 are: Use the enhanced data obtained after Step 2.2 as "positive samples" and compare the samples from different images as "negative samples". Use the following formula to compare the learning objective to optimize the feature representation: ; Among them, It is an image feature representation extracted by a neural network, is the temperature parameter, is a negative sample, N is the total number of negative samples, and the goal of this loss function is to make the feature vectors between positive samples as similar as possible, while the feature vectors of different targets are far away from each other.

[0010] Furthermore, the specific steps of Step 4 are: Step 4.1, first estimate the noise intensity of the local area in the image. Using the self-supervised learning framework, the model can automatically identify which areas in the image may contain noise, such as areas with unclear textures and blurred edges, and estimate and suppress noise based on the noise distribution model of these areas; secondly, introduce the noise weight map W to represent the noise intensity of each area. In this way, the noise area will receive a higher weight for suppression, while the target area retains more details. The noise suppression loss function is shown in the formula: ; Among them, represents the characteristics of the i-th region in the input dark spatial image S, is the denoised image The characteristics of the corresponding area in , is the adaptive noise weight adjusted according to the regional noise intensity, is the indicator function, when When it belongs to the noise area, the weight of the area increases.

[0011] Step 4.2, further introduce a self-supervised contrastive learning mechanism to perform feature learning and contrast optimization on the image after dynamic noise suppression, so as to improve the robustness and accuracy of the target features. The contrastive learning framework in self-supervised learning maximizes the similarity between different perspectives and enhanced views, so that the model can automatically learn the key features of spatial targets without labeled data, and automatically suppress the noise area. In the contrastive learning process, a mechanism that includes positive and negative sample comparison is designed. The positive samples are generated through image enhancement, and the negative samples come from different images or background areas. The goal of optimization is to make the representation of the target area in the feature space more compact, while distinguishing the background area from the target features. This optimization process can effectively reduce the impact of noise on the target features, thereby improving the detection performance of the target. The loss function of contrastive learning is shown in the formula: ; Among them, and Respectively, they are feature representations from the input image S and the enhanced image T, is the temperature coefficient, is a negative sample, and N is the number of negative samples.

[0012] ​Furthermore, the step of using the attention mechanism in Step 5 to perform feature enhancement on the key areas of the image after dynamic noise suppression in Step 4 is as follows: The potential features of the image obtained after Step 4.2 are used to enhance key areas (such as the edge and core of the target) with the help of the attention mechanism. During the training process, the model will automatically learn which areas are most important for target recognition, thereby enhancing the feature expression ability of these areas. The attention mechanism can be implemented by the following formula: ; Among them, is the learned attention weight, is the feature representation of the i-th local region.

[0013] Furthermore, the steps of using a density-based clustering algorithm method in Step 6 to classify similar features of the image after feature enhancement in Step 5 and perform target recognition and analysis are as follows: Use density-based clustering algorithm to classify similar features into the same category to identify and analyze new spatial targets. The density-based clustering algorithm can be implemented by the following formula: ; Among them, is the jth cluster center, is the sample set of the jth cluster, is a sample point in the feature representation.

[0014] Furthermore, the step of using the non-regularized self-supervised learning mechanism in Step 7 to perform image enhancement on the image analyzed in Step 6 is: Step 7.1, use bilateral filtering technology to smooth the input low-light image and preliminarily estimate the lighting map, which is achieved through the following formula: ; Among them, represents the preliminary lighting map, and S represents the input low-light image: Step 7.2 Use the self-supervised learning model to further optimize the preliminary lighting map just obtained to obtain a more accurate and consistent optimized lighting map, which can be achieved using the following formula: ; Among them, is the model parameter, obtained through optimization training; It is a self-supervised learning model; is the preliminary lighting map, and S is the input low-light image.

[0015] ​Step 7.3 Apply the correction factor r to adjust the lighting diagram to obtain the adjusted lighting diagram, which can be achieved using the following formula: ; Among them, is the reflectivity map, is the input low-light image, is the adjusted lighting diagram; The reflectivity map R is used as an enhanced image and visualized by amplifying its intensity value, which can be achieved through the following formula: ; Among them, Function used to enlarge the reflectivity map Intensity value can make the resulting image brighter and clearer.

[0016] Beneficial effects: This method can effectively enhance the target features in the image under low lighting conditions and improve the accuracy and robustness of target detection. By combining the self-supervised contrast learning mechanism and the adaptive noise suppression technology, the present invention realizes the efficient processing of low-light images and overcomes the shortcomings of traditional methods in noise suppression, contrast adjustment and feature extraction. This method does not need to rely on a large amount of labeled data, has good generalization ability and adaptability, and is suitable for high-demand application scenarios such as military reconnaissance, night monitoring and traffic monitoring, and significantly improves the detection and recognition of space targets in low-light environments. Brief Description of the Figures

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of feature extraction based on self-supervised contrastive learning; Figure 3 is a schematic diagram of spatial target feature optimization for adaptive noise suppression; Figure 4 Illustration of illumination map estimation using non-regularized self-supervised learning. Specific implementation method

[0019] ​The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Take low-light city streets as an example, such as Figure 1 As shown in the figure, the low-illumination spatial target contrast feature enhancement method based on self-supervised learning includes the following steps: Step 1, collect spatial images in low-light spatial scenes; Step 2: Perform data input and data enhancement operations on the low-light spatial image collected in Step 1, including rotation, cropping, etc. Step 3, use the self-supervised contrastive learning mechanism to extract features from the image after data enhancement in Step 2; Step 4, use adaptive noise suppression mechanism to dynamically suppress the image after feature extraction in Step 3; Step 5, use the attention mechanism to enhance the features of the key areas of the image after dynamic noise suppression in Step 4; Step 6: Use density-based clustering algorithm to classify similar features of the image enhanced in Step 5 and perform target recognition and analysis; Step7: Use the non-regularized self-supervised learning mechanism to enhance the image after the target analysis in Step6.

[0021] As a preferred implementation, the specific steps of Step 2 are: Step 2.1, perform geometric transformations including rotation, translation, cropping, scaling and flipping on the dim space image; Step 2.2, perform color space transformation including brightness adjustment, contrast adjustment, color disturbance and grayscale on the dark space image after geometric transformation, and use these enhanced samples as "positive samples", and compare the samples from different images as "negative samples".

[0022] Furthermore, the step of using the self-supervised contrastive learning mechanism in Step 3 to extract features from the image after data enhancement in Step 2 is: Use the enhanced data obtained after Step 2.2 as "positive samples" and compare the samples from different images as "negative samples". Use the following formula to compare the learning objective to optimize the feature representation: ; Among them, It is an image feature representation extracted by a neural network, is the temperature parameter, is a negative sample, N is the total number of negative samples, and the goal of this loss function is to make the feature vectors between positive samples as similar as possible, while the feature vectors of different targets are far away from each other, such as Figure 2 As shown in the experiment, the contrast and brightness of the image are significantly improved after image enhancement: the image contrast is increased by 25.43% and the brightness is improved by 15.25dB.

[0023] As a preferred implementation, the specific steps of Step 4 are: Step 4.1, first estimate the noise intensity of the local area in the image. Using the self-supervised learning framework, the model can automatically identify which areas in the image may contain noise, such as areas with unclear texture and blurred edges, and estimate and suppress noise based on the noise distribution model of these areas; secondly, introduce the noise weight map W to represent the noise intensity of each area. In this way, the noisy area will obtain a higher weight for suppression, while the target area retains more details. The noise suppression loss function is shown in the formula: ; Among them, represents the characteristics of the i-th region in the input dark spatial image S, is the denoised image The characteristics of the corresponding area in , is the adaptive noise weight adjusted according to the regional noise intensity, is the indicator function, when When it belongs to the noise area, the weight of the area increases. Experiments show that after self-supervised contrastive learning, the target recognition accuracy and feature extraction quality have been greatly improved: the target recognition accuracy has increased to 88.34%, and the feature similarity has increased by 17.89%; Step 4.2, further introduce the self-supervised contrastive learning mechanism to perform feature learning and contrast optimization on the image after dynamic noise suppression, so as to improve the robustness and accuracy of the target features. The contrastive learning framework in self-supervised learning maximizes the similarity between different perspectives and enhanced views, so that the model can automatically learn the key features of spatial targets without labeled data, and automatically suppress the noise area. In the contrastive learning process, a mechanism that includes positive and negative sample comparison is designed. The positive samples are generated by image enhancement, and the negative samples come from different images or background areas. The optimization goal is to make the representation of the target area in the feature space more compact, while distinguishing the background area from the target features. This optimization process can effectively reduce the impact of noise on the target features, thereby improving the detection performance of the target. The loss function of contrastive learning is shown in the formula: ; Among them, and Respectively, they are feature representations from the input image S and the enhanced image T, is the temperature coefficient, is a negative sample, N is the number of negative samples, such as Figure 3 As shown in the figure, in low-altitude lighting environment, the adaptive noise suppression mechanism effectively improves the signal-to-noise ratio and retains more details: SNR is increased by 16.22dB and the detail retention rate is 24.80%.

[0024] As a preferred implementation, the specific steps of Step 5 are: The potential features of the image obtained after Step 4.2 are used to enhance key areas (such as the edge and core of the target) with the help of the attention mechanism. During the training process, the model will automatically learn which areas are most important for target recognition, thereby enhancing the feature expression ability of these areas. The attention mechanism can be implemented by the following formula: ; Among them, is the learned attention weight, is the feature representation of the i-th local region.

[0025] Furthermore, the specific steps of Step 6 are: Use density-based clustering algorithm to classify similar features into the same category to identify and analyze new spatial targets. The density-based clustering algorithm can be implemented by the following formula: ; Among them, is the jth cluster center, is the sample set of the jth cluster, is a sample point in the feature representation.

[0026] As a preferred implementation, the step of using a non-regularized self-supervised learning mechanism in Step 7 to perform image enhancement on the image analyzed in Step 6 is: Step7. 1. Use bilateral filtering technology to smooth the input low-light image and preliminarily estimate the lighting map, which can be achieved by the following formula: ; Among them, represents the preliminary lighting map, S represents the input low-light image; Step 7.2, use the self-supervised learning model to further optimize the preliminary lighting map just obtained to obtain a more accurate and consistent optimized lighting map, which can be achieved using the following formula: ; Among them, is the model parameter, obtained through optimization training; It is a self-supervised learning model; is the preliminary lighting map, S is the input low-light image; Step 7.3, apply the correction factor r to adjust the lighting diagram to obtain the adjusted lighting diagram, which can be achieved using the following formula: ; Where r is the correction factor, which is automatically learned through the training process, is the adjusted lighting map, To optimize the lighting map; Step7.4, calculate the reflectivity map according to Retinex theory, which can be achieved by the following formula: ; Among them, is the reflectivity map, is the input low-light image, is the adjusted lighting diagram; The reflectivity map R is used as an enhanced image and visualized by amplifying its intensity value, which can be achieved through the following formula: ; Among them, Function used to enlarge the reflectivity map The intensity value of can make the resulting image brighter and clearer, such as Figure 4 As shown in the figure, after the image enhancement step, the brightness and contrast of the image are significantly improved: the brightness is increased by 20.21%, the contrast is increased by 25.42%, and the processing time is 5.03 seconds.

[0027] In summary, the low-lighting spatial target contrast feature enhancement method based on self-supervised learning achieves efficient enhancement and accurate detection of spatial target contrast features in low-lighting environments by introducing self-supervised learning mechanism and adaptive noise suppression technology. This method not only improves the efficiency and quality of image processing, but also enhances the robustness and accuracy of target detection, and can meet the practical application needs in various complex low-lighting environments.

[0028] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is intended to include these modifications and variations.

Claims

1. A low-light spatial target contrast feature enhancement method based on self-supervised learning, characterized in that: The steps include: Step 1: Collect spatial images in low-light spatial scenes; Step 2, perform data input and data enhancement operations on the low-illuminance spatial image collected in Step 1, including rotation and cropping; Step 3: Use the self-supervised contrastive learning mechanism to extract features from the image after data enhancement in Step 2; Step 4, using an adaptive noise suppression mechanism to dynamically suppress the image after feature extraction in Step 3; Step 5: Use the attention mechanism to enhance the features of the key areas of the image after dynamic noise suppression in Step 4; Step 6: Use density-based clustering algorithm to classify similar features of the image after feature enhancement in Step 5 and perform target recognition and analysis; Step 7: Use a non-regularized self-supervised learning mechanism to enhance the image after target analysis in Step 6.

2. The method for enhancing the contrast feature of low-illuminance space targets based on self-supervised learning according to claim 1, characterized in that: The specific steps of Step 2 are: Step 2.1, perform geometric transformations including rotation, translation, cropping, scaling and flipping on the dark space image; Step 2.2, perform color space transformation including brightness adjustment, contrast adjustment, color perturbation and grayscale on the dark space image after geometric transformation, and use these enhanced samples as "positive samples", and use samples from different images as "negative samples" for comparison.

3. The method for enhancing the contrast feature of low-illuminance space targets based on self-supervised learning according to claim 2, characterized in that: The specific steps of Step 3 are: The enhanced data obtained after Step 2.2 is used as "positive samples", and samples from different images are used as "negative samples" for comparison. The following formula is used to compare the learning objective optimization feature representation: ； in, It is the image feature representation extracted by the neural network. is the temperature parameter, are negative samples, and N is the total number of negative samples.

4. The method for enhancing the contrast feature of low-illuminance space objects based on self-supervised learning according to claim 1, characterized in that: The specific steps of Step 4 are: Step 4.

1. First, the noise intensity of the local area in the image is estimated. The self-supervised learning framework is used to automatically identify the noisy areas in the image, and the noise is estimated and suppressed based on the noise distribution model of these areas. Secondly, the noise weight map W is introduced to represent the noise intensity of each area. The noise suppression loss function is shown in the formula: ； in, represents the characteristics of the i-th region in the input dark spatial image S, is the denoised image The characteristics of the corresponding area in is the adaptive noise weight adjusted according to the regional noise intensity, is the indicator function; Step 4.2, further introduce the self-supervised contrastive learning mechanism to perform feature learning and contrast optimization on the image after dynamic noise suppression. In the contrastive learning process, a mechanism including positive and negative sample comparison is designed. The positive samples are generated by image enhancement, and the negative samples come from different images or background areas. The loss function of contrastive learning is shown in the formula: ； in, and are the feature representations from the input image S and the enhanced image T, respectively. is the temperature coefficient, are negative samples, and N is the number of negative samples.

5. The method for enhancing the contrast feature of low-illuminance space targets based on self-supervised learning according to claim 1, characterized in that: The specific steps of Step 5 are: The potential features of the image obtained after Step 4.2 are used to enhance the key areas with the help of the attention mechanism. The attention mechanism is implemented by the following formula: ； in, is the learned attention weight, is the feature representation of the i-th local region.

6. The method for enhancing the contrast feature of low-illuminance space objects based on self-supervised learning according to claim 1, characterized in that: The specific steps of Step 6 are: Using density-based clustering algorithm Clustering algorithm groups similar features into the same category. Density-based clustering algorithm is implemented by the following formula: ； in, is the jth cluster center, is the sample set of the jth cluster, is a sample point in the feature representation.

7. The method for enhancing the contrast feature of low-illuminance space objects based on self-supervised learning according to claim 1, characterized in that: The specific steps of Step 7 are: Step 7.1, use bilateral filtering technology to smooth the input low-light image and preliminarily estimate the lighting map, which is achieved by the following formula: ； in, represents the preliminary lighting map, S represents the input low-light image; Step 7.2: Use the self-supervised learning model to further optimize the preliminary lighting map just obtained, using the following formula: ； in, are model parameters obtained through optimization training; It is a self-supervised learning model; is the preliminary lighting map, S is the input low-light image; Step 7.3, apply the correction factor r to adjust the lighting diagram to obtain the adjusted lighting diagram, which is implemented using the following formula: ； Among them, r is the correction factor, which is automatically learned through the training process. is the adjusted lighting map, To optimize the lighting map; Step 7.4, calculate the reflectivity map according to the Retinex theory, which is achieved by the following formula: ； in, is the reflectivity map, is the input low-light image, is the adjusted lighting map; The reflectance map R is taken as the enhanced image and visualized by amplifying its intensity value, which is achieved by the following formula: ； in, Function used to enlarge the reflectivity map The intensity value of .

Images