An autonomous recovery method and system for unmanned aerial vehicle aerial pictures in the case of local pixel defects

Abstract

The application belongs to the technical field of image processing, and discloses a kind of for local pixel defect condition unmanned aerial vehicle aerial picture autonomous recovery method and system, comprising: using a variety of different size convolution kernel to the unmanned aerial vehicle aerial picture containing defect is carried out feature extraction, generates multiscale feature map set;Attention mechanism based neural network model analyzes multiscale feature map, locates defect area and generates mask;Repair model is constructed, based on repair model, defect area and mask, to multiscale feature map realizes end-to-end image repair;Using integrated ADCE-Net depth perception image enhancement system optimizes repair picture, outputs high-quality repair graph.The application realizes high-quality, intelligent repair, improves picture availability and analysis value, and is suitable for local pixel defect repair under a variety of complex shooting environments.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an autonomous method and system for recovering drone aerial images in cases of local pixel loss. Background Technology

[0002] In the field of drone aerial photography technology, with the continuous growth in demand for high-precision, high-resolution aerial images across various industries, drones face increasingly complex and diverse challenges when performing aerial photography missions. On the one hand, the application scenarios of drones are constantly expanding, from conventional terrain mapping and urban planning monitoring to specialized fields such as forest fire early warning, wildlife habitat monitoring, and complex geological environment exploration. These scenarios place extremely high demands on the quality and information integrity of aerial images. For example, in forest fire early warning, it is necessary to clearly and accurately identify the location and spread trend of smoke and fire sources; any pixel loss may lead to misjudgment of the fire situation. On the other hand, the interference factors encountered during drone aerial photography are becoming increasingly complex. In addition to common electromagnetic interference, cloud cover, and equipment failure, in some special environments, such as the thin atmosphere at high altitudes, the flight stability of drones is affected, resulting in shaking during image acquisition and causing local pixel blurring or loss; in severe weather conditions such as strong winds and heavy rain, the optical equipment of drones may be contaminated by moisture and dust, which can also cause local pixel anomalies.

[0003] While existing image processing techniques can repair damaged images to some extent, they still have significant limitations when dealing with localized pixel defects in drone aerial photography. Traditional interpolation-based methods, such as bilinear interpolation and bicubic interpolation, simply estimate and fill based on information from surrounding pixels, failing to effectively restore the true details of the image. When processing regions with complex textures and structures, they are prone to blurring and jagged edges, severely degrading image quality. While deep learning-based methods have made some progress in image inpainting, most models require large amounts of labeled data for training. Labeling drone aerial images is not only time-consuming and labor-intensive, but also, due to the diversity and complexity of shooting scenes, the labeled data cannot cover all possible situations, resulting in insufficient generalization ability of the models. Therefore, when faced with new and complex localized pixel defects, the restoration results are unsatisfactory.

[0004] In conclusion, existing technologies cannot effectively solve the series of problems caused by local pixel loss in drone aerial photography. There is an urgent need for an innovative and efficient method and system to meet the growing demand for drone aerial photography applications. Summary of the Invention

[0005] To address the problems of existing technologies, this invention provides an autonomous restoration method and system for drone aerial images with local pixel defects. By integrating a multi-scale convolutional neural network architecture, an attention-based neural network model, an intelligent fusion generative adversarial network (IFGAN) architecture, and a depth-sensing image enhancement system integrating brightness, contrast, color adaptive adjustment and detail enhancement network (ADCE-Net), this invention effectively overcomes the shortcomings of existing technologies in processing such images, such as low restoration accuracy, poor detail restoration, and insufficient adaptability to complex scenes. It achieves high-quality and intelligent restoration of drone aerial images with local pixel defects, significantly enhancing the usability and analytical value of the images.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] An autonomous method for recovering drone aerial images with localized pixel loss, the method comprising:

[0008] We use convolutional kernels of different sizes to extract features from drone aerial images with defects, generating a multi-scale feature map set.

[0009] A neural network model based on the attention mechanism analyzes multi-scale feature maps, locates defective regions, and generates masks.

[0010] A restoration model is constructed, and based on the restoration model, the defective region, and the mask, end-to-end image restoration is achieved on multi-scale feature maps.

[0011] The image restoration system, which integrates ADCE-Net, is used to optimize and restore the image, and output a high-quality restored image.

[0012] The preferred set of multi-scale feature maps is as follows: ,in, For detailed feature maps, This is the global structural feature map.

[0013] Preferred methods for analyzing multi-scale feature maps, locating defective regions, and generating masks using attention-based neural network models include:

[0014] The normalized multi-scale feature map is input into the attention-based neural network model, and the attention weights are calculated through a series of convolution, pooling and fully connected layer operations.

[0015] Set attention weight thresholds to identify defect areas and generate masks;

[0016] Among them, attention weight The calculation formula is:

[0017] ;

[0018] In the formula, The attention score for the target location. , and These represent the normalized feature maps at the target location, respectively. Other locations eigenvalues, The attention score for other locations.

[0019] Preferably, the restoration model adopts the Intelligent Fusion Generative Adversarial Network (IFGAN) architecture, which specifically includes a generator module and a discriminator module, and achieves end-to-end image restoration through an adversarial training mechanism.

[0020] The generator uses adaptive feature reconstruction technology and dynamic convolution kernel deformation algorithm to generate repaired images.

[0021] The discriminator distinguishes between the repaired image output by the generator and the real complete image, and optimizes the parameters of the generator and discriminator through adversarial training.

[0022] Preferred methods for optimizing and restoring images using a depth-sensing image enhancement system integrating ADCE-Net to output high-quality restored images include: ADCE-Net automatically learns image features based on a deep learning model, adjusts brightness, contrast, and color saturation, enhances details through a generative adversarial mechanism, and outputs high-quality restored images;

[0023] The methods for adjusting the brightness, contrast, and saturation of repaired images include:

[0024] ;

[0025] ;

[0026] in, At pixel position of the brightness-adjusted image The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position Brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. Contrast enhancement factor controls the degree to which dark areas of an image become darker and bright areas become brighter;

[0027] Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... :

[0028] ;

[0029] In the formula, This represents the saturation channel value after saturation enhancement. This represents the original value of the saturation channel S of the input image in the HSV color space. The saturation enhancement coefficient;

[0030] The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation, ultimately resulting in a restored image with enhanced details. .

[0031] The present invention also provides an autonomous recovery system for UAV aerial images with local pixel defects. The system is used to implement the aforementioned method and includes: an image feature extraction module, a defect area localization module, a repair model construction module, and a repair result optimization module.

[0032] The image feature extraction module is used to extract features from drone aerial images with defects using convolution kernels of various sizes, and generate a multi-scale feature map set.

[0033] The defect region localization module is used to analyze multi-scale feature maps based on an attention mechanism neural network model, locate the defect region, and generate a mask.

[0034] The repair model construction module is used to construct a repair model, and based on the repair model, the defective region, and the mask, to achieve end-to-end image repair on multi-scale feature maps;

[0035] The repair result optimization module is used to optimize and repair images using a depth-sensing image enhancement system integrating ADCE-Net, and output high-quality repaired images.

[0036] The preferred set of multi-scale feature maps is as follows: ,in, For detailed feature maps, This is the global structural feature map.

[0037] Preferably, the defect area localization module includes: a calculation unit and a comparison unit;

[0038] The computing unit is used to input the normalized multi-scale feature map into the attention-based neural network model, and calculate the attention weights through a series of convolution, pooling and fully connected layer operations.

[0039] The comparison unit is used to set an attention weight threshold to determine the defect area and generate a mask;

[0040] Among them, attention weight The calculation formula is:

[0041] ;

[0042] In the formula, The attention score for the target location. , and These represent the normalized feature maps at the target location, respectively. Other locations eigenvalues, The attention score for other locations.

[0043] Preferably, the restoration model adopts the Intelligent Fusion Generative Adversarial Network (IFGAN) architecture, which specifically includes a generator module and a discriminator module, and achieves end-to-end image restoration through an adversarial training mechanism.

[0044] The generator uses adaptive feature reconstruction technology and dynamic convolution kernel deformation algorithm to generate repaired images.

[0045] The discriminator distinguishes between the repaired image output by the generator and the real complete image, and optimizes the parameters of the generator and discriminator through adversarial training.

[0046] Preferably, the process of optimizing and restoring images using a depth-sensing image enhancement system integrating ADCE-Net and outputting high-quality restored images includes: ADCE-Net automatically learns image features based on a deep learning model, adjusts brightness, contrast and color saturation, enhances details through a generative adversarial mechanism, and outputs high-quality restored images;

[0047] The process of adjusting the brightness, contrast, and saturation of the restored image includes:

[0048] ;

[0049] ;

[0050] in, At pixel position of the brightness-adjusted image The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position Brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. Contrast enhancement factor controls the degree to which dark areas of an image become darker and bright areas become brighter;

[0051] Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... :

[0052] ;

[0053] In the formula, This represents the saturation channel value after saturation enhancement. This represents the original value of the saturation channel S of the input image in the HSV color space. The saturation enhancement coefficient;

[0054] The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation, ultimately resulting in a restored image with enhanced details. .

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] Significantly Improved Detail Restoration and Repair Accuracy: This invention captures detailed features using small 3×3 convolutional kernels and obtains global structural features using large 7×7 convolutional kernels, providing accurate feature information for repair. Simultaneously, the generator in the IFGAN architecture utilizes adaptive feature reconstruction technology and a dynamic convolutional kernel deformation algorithm to generate repaired content that closely matches the original image based on multi-scale feature maps and masked loss regions. The discriminator distinguishes between the repaired image output by the generator and the original complete image. Adversarial training optimizes the parameters of the generator and discriminator, greatly improving repair accuracy, effectively restoring image details, and making the repaired image clearer and more realistic.

[0057] High adaptability to complex scenarios and excellent generalization ability: This invention uses an attention-based neural network model to locate missing regions. This model introduces residual connections to avoid gradient vanishing, enabling it to learn and locate missing regions more accurately under various complex conditions. Simultaneously, the entire autonomous restoration system supports online updating of model parameters. It can fine-tune the models of each module by collecting new aerial image data, enhancing its adaptability to different shooting environments, scene changes, and various local pixel defects. It exhibits stronger generalization ability and can stably and efficiently handle image restoration tasks in a variety of complex scenarios. Attached Figure Description

[0058] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of a method for autonomous recovery of drone aerial images in the case of partial pixel loss, according to an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] Example 1

[0063] This embodiment provides an autonomous recovery method for drone aerial images with local pixel loss, referring to... Figure 1 The steps are as follows: Use convolutional kernels of different sizes to extract features from drone aerial images containing defects, and generate a multi-scale feature map set;

[0064] A neural network model based on the attention mechanism analyzes multi-scale feature maps, locates defective regions, and generates masks.

[0065] A restoration model is constructed, and based on the restoration model, the defective region, and the mask, end-to-end image restoration is achieved on multi-scale feature maps.

[0066] The image restoration system, which integrates ADCE-Net, is used to optimize and restore the image, and output a high-quality restored image.

[0067] The specific implementation process is as follows:

[0068] Step 100 begins with data input, using drone aerial images with partial pixel defects as the starting data. These images are affected by factors such as electromagnetic interference and equipment malfunction, resulting in the loss or damage of some pixel information.

[0069] Step 101: In the image feature extraction stage, a differentiated scale convolutional collaborative extraction mechanism is adopted: a 3×3 small convolutional kernel is used to specifically capture local subtle features, while a 7×7 large convolutional kernel focuses on mining global structural features, thereby generating a multi-scale feature map set. The convolution operation formula for the small convolutional kernel is:

[0070] ;

[0071] in, It is the result of convolution with small kernels at the target location. eigenvalues, Is the small convolution kernel in other locations? weight , Indicates the input image in The pixel values. Large kernel convolution operations are similar:

[0072] ;

[0073] in, and These represent the feature values ​​and weights after convolution with a large convolution kernel, respectively. Through these two convolution operations, detailed feature maps are obtained. and global structural feature map This leads to the formation of a multi-scale feature map set. .

[0074] Step 102: To improve the stability and efficiency of subsequent processing, the multi-scale feature maps are normalized using the following formula:

[0075] ;

[0076] in, These are the normalized eigenvalues. and These are the minimum and maximum values ​​in the multi-scale feature map set F, respectively. This step lays the foundation for subsequent accurate analysis of image features, and its output, the normalized multi-scale feature map, is a key input for localizing the defective region.

[0077] Step 200: In the defect region localization step, the normalized multi-scale feature map is... The input is an attention-based neural network model. This model consists of three main modules: a feature encoding module, an attention scoring module, and a weight normalization module. Its core objective is to achieve accurate localization of defective regions in an image, providing crucial spatial guidance information for subsequent image inpainting. The specific structure is as follows:

[0078] First, normalize the input features. Channel compression and feature abstraction are performed using two stacked convolutional layers:

[0079] ;

[0080] ;

[0081] Here, Q and K are the query and key vectors, respectively, representing the role of each pixel position in the attention mechanism.

[0082] Secondly, a dot product-based attention mechanism is used to score the position of each pixel in the feature map. and These represent the normalized feature maps at the target location, respectively. Other locations The eigenvalues, hence the attention score The calculation is as follows:

[0083] .

[0084] Finally, the scoring results are normalized using the Softmax function to obtain the attention weights at each pixel location. The calculation formula is as follows:

[0085] .

[0086] Step 201, set the attention weight threshold T=0.8, when When the area is identified as a defective region, a defective region mask is generated. Where H represents the image height and W represents the image width. The mask size is the same as the input image, and the target location is defined in the defective region. superior Other locations This indicates that the image structure in this region is intact. The specific formula is as follows:

[0087] ;

[0088] This step locates the defective region based on the feature map information, providing crucial location information for the repair model. The output defective region mask is one of the important inputs to the generator in the repair model construction.

[0089] Step 300: The restoration model is constructed using an Intelligent Fusion Generative Adversarial Network (IFGAN) architecture. Specifically, it includes a generator module and a discriminator module, achieving end-to-end image restoration through an adversarial training mechanism. The generator uses adaptive feature reconstruction technology and dynamic convolutional kernel deformation algorithms to generate restored images; the discriminator distinguishes between the restored images output by the generator and the original complete images, optimizing the parameters of the generator and discriminator through adversarial training.

[0090] The generator uses a normalized set of multi-scale feature maps and a mask of the missing regions. During the feature decoding stage, the generator introduces an adaptive feature reconstruction technique, which involves constructing a lightweight feature guidance module to dynamically weight and fuse feature maps of different scales and semantic levels, thereby improving the ability to recover details from the missing regions. The fusion method is shown below:

[0091] ;

[0092] in, For the number of feature layers, This represents the output fused feature map. This represents the feature map of the nth layer. Indicates the feature map Upsampling is performed to unify the resolution. Furthermore, , representing the normalization constraint on the weight coefficients. The above operations map multi-scale feature maps from low-dimensional space to high-dimensional image space through multi-layer transposed convolution and upsampling operations. At the same time, to improve structural continuity and texture restoration accuracy, the generator introduces skip connections to fuse shallow detail features and deep semantic features, avoiding semantic breaks between the repaired area and the original image.

[0093] Step 301: The multi-scale feature map set and the defect region mask are input into the dynamic convolutional kernel deformation algorithm. The shape of the convolutional kernel, the spatial weight distribution, and the inter-channel correlation are adjusted in real time through the convolutional kernel parameter prediction network. The formula is:

[0094] ;

[0095] Where G represents the repaired feature map, K is the dynamic convolution kernel, F represents the input feature map, and M represents the mask. For dynamic convolution kernel parameters, This indicates element-wise multiplication, ensuring that multiplication occurs only in the missing regions. Perform repair feature generation.

[0096] Step 302: Use binary cross-entropy (BCE) to perform adversarial optimization on the discriminator loss function, with the goal of maximizing the discrimination probability of real images. and minimizing the discrimination probability of the repaired image ,Right now

[0097] ;

[0098] In the formula, The loss function of the discriminator measures its ability to distinguish between real and generated images. This represents the expectation for real image samples and evaluates the model's performance on real images. This represents the expectation for the generated image samples and evaluates the model's performance on the generated images.

[0099] Generator loss function To minimize the discriminant's probability of classifying the repaired image, the following formula is used:

[0100] ;

[0101] An alternating optimization strategy is employed (the discriminator parameters are updated first, then the generator parameters in each training round), using the Adam optimizer (learning rate). Iterate through the training until the loss function converges.

[0102] Step 400: In the restoration result optimization stage, the restoration image is comprehensively optimized using a depth-aware image enhancement system (DAIE) integrating ADCE-Net (Adaptive adjustment of brightness, contrast, color and details Enhancement Network). Regarding ADCE-Net adaptive adjustment, the brightness, contrast, and saturation of the restoration image are adjusted using the following formula:

[0103] ;

[0104] ;

[0105] in, At pixel position of the brightness-adjusted image The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position Brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. The contrast enhancement factor controls how much dark areas of an image become darker and bright areas become brighter.

[0106] Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... :

[0107] ;

[0108] In the formula, This represents the saturation channel value after saturation enhancement. This represents the original value of the saturation channel S of the input image in the HSV color space. This is the saturation enhancement factor. The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation to finally obtain the restored image with enhanced details. .

[0109] In summary, the system of this invention proceeds sequentially through each step, with the output of the previous step serving as the input for the next. The system includes modules for image feature extraction, defect region localization, repair model construction, and repair result optimization. These modules work together to achieve efficient repair of drone aerial images with local pixel defects. This effectively solves the problems of low repair accuracy, poor detail restoration, and insufficient adaptability to complex scenes in existing technologies when processing drone aerial images with local pixel defects. It achieves high-quality, intelligent repair, improves image usability and analytical value, and is suitable for repairing local pixel defects in various complex shooting environments, meeting practical application needs.

[0110] Example 2

[0111] The present invention also provides an autonomous recovery system for UAV aerial images with local pixel defects. The system is used to implement the method described in Embodiment 1. The system includes: an image feature extraction module, a defect area localization module, a repair model construction module, and a repair result optimization module.

[0112] The image feature extraction module is used to extract features from drone aerial images with defects using convolution kernels of various sizes, generating a multi-scale feature map set.

[0113] The defect region localization module is used to analyze multi-scale feature maps using a neural network model based on an attention mechanism, locate the defect region, and generate a mask.

[0114] The repair model construction module is used to build a repair model. Based on the repair model, the defective region, and the mask, it can achieve end-to-end image repair on multi-scale feature maps.

[0115] The restoration result optimization module is used to optimize and restore images using a depth-aware image enhancement system integrated with ADCE-Net, and output high-quality restored images.

[0116] In this embodiment, the multi-scale feature map set is as follows: ,in, For detailed feature maps, This is the global structural feature map.

[0117] In this embodiment, the defect area localization module includes: a calculation unit and a comparison unit;

[0118] The computation unit is used to input the normalized multi-scale feature map into the attention-based neural network model and calculate the attention weights through a series of convolution, pooling and fully connected layer operations.

[0119] The comparison unit is used to set the attention weight threshold, determine the defect area, and generate a mask;

[0120] Among them, attention weight The calculation formula is:

[0121] ;

[0122] In the formula, The attention score for the target location. , and These represent the normalized feature maps at the target location, respectively. Other locations eigenvalues, The attention score for other locations.

[0123] In this embodiment, the restoration model adopts the Intelligent Fusion Generative Adversarial Network (IFGAN) architecture, which specifically includes a generator module and a discriminator module. End-to-end image restoration is achieved through an adversarial training mechanism.

[0124] The generator uses adaptive feature reconstruction technology and dynamic convolution kernel deformation algorithm to generate repaired images.

[0125] The discriminator distinguishes between the repaired image output by the generator and the real complete image, and optimizes the parameters of the generator and discriminator through adversarial training.

[0126] In this embodiment, the process of optimizing and repairing images using a depth-sensing image enhancement system integrating ADCE-Net and outputting high-quality repaired images includes: ADCE-Net automatically learns image features based on a deep learning model, adjusts brightness, contrast and color saturation, enhances details through a generative adversarial mechanism, and outputs high-quality repaired images;

[0127] The process of adjusting the brightness, contrast, and saturation of the restored image includes:

[0128] ;

[0129] ;

[0130] in, At pixel position of the brightness-adjusted image The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position Brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. Contrast enhancement factor controls the degree to which dark areas of an image become darker and bright areas become brighter;

[0131] Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... :

[0132] ;

[0133] In the formula, This represents the saturation channel value after saturation enhancement. This represents the original value of the saturation channel S of the input image in the HSV color space. The saturation enhancement coefficient;

[0134] The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation, ultimately resulting in a restored image with enhanced details. .

[0135] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for autonomous recovery of drone aerial images with local pixel loss, characterized in that, The method includes: We use convolutional kernels of different sizes to extract features from drone aerial images with defects, generating a multi-scale feature map set. A neural network model based on the attention mechanism analyzes multi-scale feature maps, locates defective regions, and generates masks. A restoration model is constructed, and based on the restoration model, the defective region, and the mask, end-to-end image restoration is achieved on multi-scale feature maps. The image restoration system, which integrates ADCE-Net, is used to optimize and restore the image, and output a high-quality restored image. The restoration model adopts the intelligent fusion generative adversarial network architecture IFGAN, which specifically includes a generator module and a discriminator module. It achieves end-to-end image restoration through an adversarial training mechanism. The generator uses adaptive feature reconstruction technology and dynamic convolution kernel deformation algorithm to generate repaired images. The discriminator distinguishes between the repaired image output by the generator and the real complete image, and optimizes the parameters of the generator and discriminator through adversarial training; Methods for optimizing and restoring images using a deep-aware image enhancement system integrating ADCE-Net to output high-quality restored images include: ADCE-Net automatically learns image features based on a deep learning model, adjusts brightness, contrast and color saturation, enhances details through a generative adversarial mechanism, and outputs high-quality restored images; The methods for adjusting the brightness, contrast, and saturation of repaired images include: ； ； in, At pixel position of the image after brightness adjustment The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position The brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. Contrast enhancement factor controls the degree to which dark areas of an image become darker and bright areas become brighter; Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... : ； In the formula, This represents the saturation channel value after saturation enhancement. This represents the saturation channel of the input image in the HSV color space. S The original value, This is the saturation enhancement coefficient; The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation, ultimately resulting in a restored image with enhanced details. .

2. The method according to claim 1, characterized in that, The multi-scale feature map set is as follows: ,in, For detailed feature maps, This is the global structural feature map.

3. The method according to claim 1, characterized in that, Methods for analyzing multi-scale feature maps using attention-based neural network models to locate defect regions and generate masks include: The normalized multi-scale feature map is input into the attention-based neural network model, and the attention weights are calculated through a series of convolution, pooling and fully connected layer operations. Set attention weight thresholds to identify defect areas and generate masks; Among them, attention weight The calculation formula is: ； In the formula, The attention score for the target location. , and These represent the normalized feature maps at the target location. Other locations eigenvalues, The attention score for other locations.

4. An autonomous recovery system for UAV aerial images with localized pixel loss, the system being used to implement the method described in any one of claims 1-3, characterized in that, The system includes: an image feature extraction module, a defect area localization module, a repair model construction module, and a repair result optimization module; The image feature extraction module is used to extract features from drone aerial images with defects using convolution kernels of various sizes, and generate a multi-scale feature map set. The defect region localization module is used to analyze multi-scale feature maps based on an attention mechanism neural network model, locate the defect region, and generate a mask. The repair model construction module is used to construct a repair model, and based on the repair model, the defective region, and the mask, to achieve end-to-end image repair on multi-scale feature maps; The repair result optimization module is used to optimize and repair images using a depth-sensing image enhancement system integrating ADCE-Net, and output high-quality repaired images.

5. The system according to claim 4, characterized in that, The multi-scale feature map set is as follows: ,in, For detailed feature maps, This is the global structural feature map.

6. The system according to claim 4, characterized in that, The defect area localization module includes: a calculation unit and a comparison unit; The computing unit is used to input the normalized multi-scale feature map into the attention-based neural network model and calculate the attention weights through a series of convolution, pooling and fully connected layer operations. The comparison unit is used to set an attention weight threshold to determine the defect area and generate a mask; Among them, attention weight The calculation formula is: ； In the formula, The attention score for the target location. , and These represent the normalized feature maps at the target location. Other locations eigenvalues, The attention score for other locations.

7. The system according to claim 4, characterized in that, The restoration model adopts the intelligent fusion generative adversarial network architecture IFGAN, which specifically includes a generator module and a discriminator module. It achieves end-to-end image restoration through an adversarial training mechanism. The generator uses adaptive feature reconstruction technology and dynamic convolution kernel deformation algorithm to generate repaired images. The discriminator distinguishes between the repaired image output by the generator and the real complete image, and optimizes the parameters of the generator and discriminator through adversarial training.

8. The system according to claim 4, characterized in that, The process of optimizing and restoring images using the integrated ADCE-Net deep-aware image enhancement system and outputting high-quality restored images includes: ADCE-Net automatically learns image features based on a deep learning model, adjusts brightness, contrast and color saturation, enhances details through generative adversarial mechanisms, and outputs high-quality restored images; The process of adjusting the brightness, contrast, and saturation of the restored image includes: ； ； in, At pixel position of the image after brightness adjustment The brightness value at that location, To initially repair the image at pixel locations The original brightness value at that location, This is a brightness scaling factor used to adjust the overall brightness of the image. Brightness offset is used to fine-tune the overall brightness increase. Image with contrast adjusted at pixel position The brightness value; This is the average brightness value of the image, used for enhancement based on the brightness of the image center. Contrast enhancement factor controls the degree to which dark areas of an image become darker and bright areas become brighter; Color saturation adjustment is performed in the HSV color space. In the HSV space, the saturation channel is S, and the adjusted saturation channel is... : ； In the formula, This represents the saturation channel value after saturation enhancement. This represents the saturation channel of the input image in the HSV color space. S The original value, This is the saturation enhancement coefficient; The adjusted saturation channel is merged with other channels to obtain the adjusted image, and the generated details are continuously optimized through backpropagation, ultimately resulting in a restored image with enhanced details. .

Images