Image enhancement method suitable for monitoring unmanned aerial vehicle

Abstract

The invention discloses an image enhancement method suitable for monitoring an unmanned aerial vehicle, and relates to the technical field of unmanned aerial vehicles. The method comprises the steps of estimating and inputting an adaptive illumination enhancement high-order curve through a depth illumination curve estimation network, adjusting the brightness and details of low-light and high-light areas in real time through the adaptive illumination enhancement high-order curve, ensuring that a target in an image is always clear and visible, and improving the image quality. The loss function of the depth illumination curve estimation network is the sum of a reference-free image loss function and a content loss function; the content loss function is used for generating the pixel-level difference of the image and the target image in the corresponding layer through calculation, and taking the result as a loss value. According to the invention, through the adaptive illumination enhancement high-order curve algorithm, each pixel is subjected to enhancement of different intensities according to the illumination condition of the position where the pixel is located, through enhancement of a local area, the algorithm can recover details of a shielded area, the accuracy of target detection and tracking is improved, and stable tracking of a target is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles, and particularly relates to an image enhancement method applicable to surveillance unmanned aerial vehicles. Background Art

[0002] When a surveillance unmanned aerial vehicle performs outdoor high-altitude operations, the unmanned aerial vehicle will enter various different environments and geographical regions during flight, and the lighting conditions will change with the geographical location, time, and weather. Therefore, one of the main technical challenges faced by the unmanned aerial vehicle is how to maintain the stability of the shooting quality of the surveillance camera under such complex lighting conditions, so as to improve the task execution efficiency. Complex lighting conditions usually include two situations: backlight and strong light. Especially in the backlight environment, uneven lighting often leads to unbalanced brightness and darkness of the image, and phenomena such as overexposure or increased shadows often occur in the image, making it difficult to distinguish the target from the background, and even causing the target to merge with the background, resulting in difficulty in extracting effective features. Under strong light irradiation, there will also be uneven brightness and darkness in the image. The shadow area may block the target features, affecting the presentation of image details, and even causing the target contour to be blurred. These problems make the traditional target detection algorithms perform unstably in environments with poor lighting, and the recognition accuracy drops significantly.

[0003] Currently, there are some image enhancement methods based on deep learning, such as the multi-scale Retinex (MSR) algorithm and the EnlightenGAN algorithm. These methods have improved the image quality in low-light and backlight environments to a certain extent. The multi-scale Retinex algorithm decomposes the image into two parts: brightness and color, and processes it at multiple scales, which can enhance the local details and global features of the image, and improve the contrast and brightness. However, since this method performs weighted averaging between different scales, especially when processing images with rich details, artifacts or noise may be introduced.

[0004] The EnlightenGAN algorithm is based on the generative adversarial network (GAN), and can perform adaptive image enhancement without paired training data. By introducing an attention mechanism and a self-feature retention loss, this algorithm improves the quality and authenticity of the generated image, and performs well especially under complex lighting conditions. However, due to its dependence on adversarial training, the quality of the generated image may be unstable, and especially under extreme lighting conditions, it may still have a negative impact on the enhancement effect.

[0005] Although these methods have improved the image quality to some extent, they are insufficient in stability in complex environments, have a slow processing speed, require a large amount of computing resources, and also have certain deficiencies in dealing with strong light and shadows. Summary of the Invention

[0006] The object of the present invention is to provide an image enhancement method applicable to monitoring drones. Through the adaptive illumination enhancement high-order curve algorithm, each pixel is enhanced with different intensities according to the illumination conditions at its location. Through the enhancement of the local area, the algorithm can restore the details of the occluded area, make the target and the background more distinct, improve the accuracy of target detection and tracking, ensure the stable tracking of the target, and solve the problem of insufficient stability of image enhancement in the existing complex environment.

[0007] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0008] The present invention is an image enhancement method applicable to monitoring drones, including: estimating the input adaptive illumination enhancement high-order curve through a depth illumination curve estimation network, and through the adaptive illumination enhancement high-order curve, adjusting the brightness and details of the low-light and high-light regions in real time to ensure that the target in the image is always clearly visible, and obtaining an enhanced image; the adaptive illumination enhancement high-order curve is:

[0009] E n (x,y) = E n-1 (x,y) + A n E n-1 (x,y)(1 - E n-1 (x,y))

[0010] Among them, (x,y) is the pixel coordinate in the image, E n (x,y) is the result after the image is enhanced, n represents the number of iterations, A n is a parameter matrix, the size of which is the same as that of the input image; the loss function of the depth illumination curve estimation network is the sum of the non-reference image loss function and the content loss function; the content loss function is used to calculate the pixel-level difference between the generated image and the target image at the corresponding layer, and take the result as the loss value; the non-reference image loss function is the sum of the spatial consistency loss, the exposure control loss, the color constancy loss and the illumination smoothness loss; the spatial consistency loss is to maintain the spatial consistency of the image by restricting the difference between adjacent regions of the input image and the enhanced image; the exposure control loss controls the distance between the intensity mean of the local area of the enhanced image and the ideal exposure level, so as to suppress the exposure degree of the image; the color constancy loss controls the color deviation between the enhanced image and the input image by establishing the relationship between the three RGB channels; the illumination smoothness loss restricts the mean of the horizontal and vertical gradients in all channels and the number of iterations to maintain the monotonic relationship between adjacent pixels.

[0011] As a preferred technical solution of the present invention, the depth illumination curve estimation network estimates the enhancement curve of the input image through a convolutional neural network.

[0012] As a preferred technical solution of the present invention, the deep illumination curve estimation network consists of seven symmetrically cascaded convolutional layers. The first six layers are composed of convolution Conv and ReLU activation functions for extracting image features. Among them, the outputs of the first three convolutional layers are cascaded with the outputs of the last three convolutional layers at the channel level to make full use of feature information at different levels. The last layer uses the tanh activation function and finally outputs a set of pixel-level curve parameters describing the corresponding high-order brightness curve.

[0013] As a preferred technical solution of the present invention, each convolutional layer uses a 3×3×32 convolutional kernel.

[0014] As a preferred technical solution of the present invention, the calculation method of the spatial consistency loss is as follows:

[0015]

[0016] Among them, K is the number of local regions, Ω(i) represents four adjacent regions centered on the region, and Y and I are the intensity means of the local regions of the enhanced image and the input image, respectively.

[0017] As a preferred technical solution of the present invention, the calculation method of the exposure control loss is as follows:

[0018]

[0019] Among them, H takes 0.6, and the formula is as follows, where M is the number of local regions, with a size of 16×16, and Y k is the intensity mean of the local region of the enhanced image.

[0020] As a preferred technical solution of the present invention, the calculation method of the color constancy loss is as follows:

[0021]

[0022] Among them, where J p represents the intensity mean of the p-th channel of the enhanced image, and (p, q) represents the paired channels of the enhanced image and the input image.

[0023] As a preferred technical solution of the present invention, the calculation method of the illumination smoothness loss is as follows:

[0024]

[0025] Among them, N is the number of iterations, and respectively represent the gradient calculations in the horizontal and vertical directions. As a preferred technical solution of the present invention, the loss function without a reference image is:

[0026] L total = L spa + L exp + w col L col + w tvA L tvA

[0027] Wherein, w col and w tvA are the weights of the color constancy loss and the illumination smoothness loss, respectively.

[0028] As a preferred technical solution of the present invention, the calculation formula of the content loss function is:

[0029]

[0030] Wherein, represents the feature map extracted from the i-th layer of the VGG16 network, I is the input image, G is the generated image, and C i , H i , W i are the number of channels, height, and width of the i-th layer feature map, respectively.

[0031] The present invention has the following beneficial effects:

[0032] The present invention estimates the input adaptive illumination enhancement high-order curve through the depth illumination curve estimation network. Through the adaptive illumination enhancement high-order curve, especially in the face of strong light or backlight environments, the algorithm can continuously optimize the adaptive illumination enhancement high-order curve, and can adjust the brightness and contrast of the image in real time under dynamic illumination conditions, so as to maintain the clarity and recognizability of the target in the image.

[0033] Moreover, by introducing the adaptive enhancement matrix, each pixel can be enhanced with different intensities according to the illumination conditions at its position. Through the enhancement of the local area, the algorithm can restore the details of the occluded area, making the target more distinct from the background. Especially in the case of partial occlusion of the target or complex background, the accuracy of target detection and tracking is improved, ensuring the stable tracking of the target.

[0034] Of course, it is not necessary for any product implementing the present invention to achieve all of the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 Schematic diagram of the depth illumination curve estimation network;

[0037] Figure 2 Schematic diagram of the VGG16 network. Specific implementation manners

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] The present invention is an image enhancement method applicable to monitoring unmanned aerial vehicles, including: estimating an input adaptive illumination enhancement high-order curve through a depth illumination curve estimation network, and adjusting the brightness and details of low-light and high-light regions in real time through the adaptive illumination enhancement high-order curve to ensure that the targets in the image are always clearly visible, thereby obtaining an enhanced image.

[0040] Specifically, the iterative enhancement algorithm based on the image enhancement matrix transforms the image enhancement problem into a process of optimizing the illumination enhancement curve. To ensure the effect and performance of the algorithm, the illumination enhancement curve should meet the requirements of simplicity, monotonic differentiability, and normalization when designed. Monotonic differentiability ensures the preservation of the correlation between pixels at different positions in the image, while normalization can prevent the loss of image information caused by truncation. The representation form of the illumination enhancement curve is as follows:

[0041] E(I(x,y); α) = I(x,y) + αI(x,y)(1 - I(x,y))

[0042] Among them, (x,y) are the pixel coordinates in the image, E(I(x,y); α) is the result after the image is enhanced, α ∈ [-1,1] is an adjustable parameter used to control the amplitude of the enhancement curve. This curve can be applied to the RGB three channels respectively to ensure that the color of the image is not affected and overfitting is avoided as much as possible. To adapt to different illumination conditions, the algorithm continuously improves the estimation effect of the curve by iteratively optimizing the curve mapping. Especially in backlit images taken by unmanned aerial vehicles, this method can significantly improve the details and visibility of the image, thereby improving the accuracy of subsequent target detection and recognition. Its formula is:

[0043] E n (x,y) = E n-1 (x,y) + α n E n-1 (x,y)(1 - E n-1 (x,y))

[0044] Among them, n represents the number of iterations. Compared with the first-order illumination enhancement curve, the high-order curve has a stronger adjustment ability and can process complex illumination environments more precisely. However, since α n is a global parameter, it can only enhance the image globally. Therefore, the algorithm introduces the enhancement parameter matrix A n , whose size is the same as that of the input image, and can perform adaptive enhancement processing on each pixel. In this way, the algorithm can adaptively adjust the enhancement effect of each pixel according to the specific requirements of the image, thereby improving the accuracy and effect of enhancement. At this time, the high-order curve of adaptive illumination enhancement is:

[0045] E n (x, y) = E n-1 (x, y) + A n E n-1 (x, y)(1 - E n-1 (x, y))

[0046] Among them, (x, y) are the pixel coordinates in the image, E n (x, y) is the result of the image after enhancement, n represents the number of iterations, and A n is the parameter matrix, whose size is the same as that of the input image.

[0047] In the UAV inspection task, especially when facing strong light or backlight environments, the algorithm continuously optimizes the illumination enhancement curve to ensure that the targets in the image are always clearly visible and avoid target loss caused by backlight or strong light. This adaptive enhancement curve method can adjust the brightness and contrast of the image in real time under dynamic illumination conditions, thereby maintaining the clarity and recognizability of the targets in the image.

[0048] In addition, by introducing the adaptive enhancement parameter matrix, the algorithm enables each pixel to be enhanced with different intensities according to the illumination conditions at its location. Through local area enhancement, the algorithm can restore the details of the occluded areas, making the targets more distinct from the background. Especially in the case of partial occlusion of the targets or complex backgrounds, it improves the accuracy of target detection and tracking and ensures the stable tracking of the targets.

[0049] As Figure 1 shown, the loss function of the deep illumination curve estimation network is the sum of the no-reference image loss function and the content loss function. The deep illumination curve estimation network estimates the enhancement curve of the input image through a convolutional neural network. The deep illumination curve estimation network consists of seven symmetrically cascaded convolutional layers, where the first six layers consist of a convolution Conv and a ReLU activation function, and are used to extract the features of the image.

[0050] Among them, the outputs of the first three convolutional layers are concatenated with those of the last three convolutional layers at the channel level to make full use of feature information at different levels; the last layer uses the tanh activation function, and finally outputs a set of pixel-level curve parameters describing the corresponding high-order brightness curve. Each convolutional layer uses a 3×3×32 convolutional kernel. During the training process, this algorithm optimizes the mapping relationship between the original image and the enhanced image as the objective function. Through training, the network can learn the complex mapping relationship between the input image and its enhancement curve, so as to effectively enhance the input image.

[0051] The loss function without a reference image is the sum of the spatial consistency loss, the exposure control loss, the color constancy loss, and the illumination smoothness loss.

[0052] Specifically, the spatial consistency loss maintains the spatial consistency of the image by restricting the difference between adjacent regions of the input image and the enhanced image; the calculation method of the spatial consistency loss is:

[0053]

[0054] where K is the number of local regions, Ω(i) represents the four adjacent regions centered on the region, and Y and I are the intensity means of the local regions of the enhanced image and the input image, respectively.

[0055] The exposure control loss suppresses the exposure degree of the image by controlling the distance between the intensity mean of the local region of the enhanced image and the ideal exposure level; the calculation method of the exposure control loss is:

[0056]

[0057] where H is taken as 0.6, and the formula is as follows, where M is the number of local regions, with a size of 16×16, and Y k is the intensity mean of the local region of the enhanced image.

[0058] The color constancy loss controls the color deviation between the enhanced image and the input image by establishing the relationship between the three RGB channels; the calculation method of the color constancy loss is:

[0059]

[0060] where J p represents the intensity mean of channel p of the enhanced image, and (p,q) represents the paired channels of the enhanced image and the input image.

[0061] The illumination smoothness loss restricts the mean of the horizontal and vertical gradients in all channels and the number of iterations to maintain the monotonic relationship between adjacent pixels. The calculation method of the illumination smoothness loss is:

[0062]

[0063] Among them, N is the number of iterations, and respectively represent the gradient calculations in the horizontal and vertical directions.

[0064] Therefore, the loss function without a reference image is:

[0065] L total = L spa + L exp + w col L col + w tvA L tvA

[0066] Among them, w col and w tvA are the weights of the color constancy loss and the illumination smoothness loss respectively.

[0067] Since the depth illumination curve estimation network consists of seven convolutional layers, the feature learning ability of the network is enhanced by the way of multi-layer symmetric cascading. Each layer of convolutional operation may introduce a certain amount of detail loss. Although skip connections are added to retain the underlying information, the image may still lose some details during the multi-layer convolutional process. The content loss function is used to calculate the pixel-level difference between the generated image and the target image at the corresponding layer and take the result as the loss value.

[0068] Therefore, in order to retain the details of the image as much as possible, this algorithm introduces a content loss function to balance the detail differences between the generated image and the original image. The content loss function is used in deep learning to measure the similarity between the generated image and the target image. It extracts image features through a pre-trained convolutional neural network and compares the feature representations of the generated image and the target image at a certain layer.

[0069] As Figure 2 shown, taking the selection of the VGG16 network to extract image features as an example. The VGG16 network consists of 16 layers of neural networks, mainly including convolutional layers and pooling layers, and outputs results through the last few fully connected layers. The deep convolutional layers and smaller convolutional kernels of VGG16 improve the network's ability in feature representation and non-linear modeling. Since the content loss is insensitive to image intensity changes, it can effectively limit the content differences between the generated image and the previous enhanced image and help retain more detail information

[0070] The calculation formula of the content loss function is:

[0071]

[0072] Among them, denotes the feature map extracted from the i-th layer of the VGG16 network, where I is the input image, G is the generated image, and C i , H i , W i are the number of channels, height, and width of the feature map of the i-th layer, respectively.

[0073] By calculating the Euclidean distance between the feature maps of the generated image and the previous enhanced image, the content loss function helps the algorithm judge the differences before and after enhancement, ensuring that the details of the image are retained in each iteration. Finally, the total loss function of the algorithm combines spatial consistency, exposure control, color constancy, illumination smoothing, and content loss, and the formula is as follows:

[0074] L total = L spa + L exp + w col L col + w tvA L tvA + L content

[0075] Through this combined loss function, this algorithm can effectively optimize the image enhancement process, retain the detail information of the image, and improve the quality and consistency of the image.

[0076] This algorithm aims to solve the problem of image quality degradation of drones in complex lighting environments, such as backlight environments, and has the following significant advantages:

[0077] Precise and real-time image enhancement: This algorithm combines image matrix enhancement technology to enhance the low-light and high-light regions in the image in real time, ensuring that the target is always clearly visible in complex lighting environments. Especially in backlight environments, the algorithm effectively restores the occluded areas or dark details, maintaining the high visibility of the target, thus improving the image quality.

[0078] Adaptive enhancement adjustment: This algorithm introduces an adaptive enhancement matrix to dynamically adjust according to the lighting conditions of each pixel. This mechanism can automatically optimize the image brightness, avoid overexposure or underexposure, and thus significantly improve the overall quality and detail performance of the image.

[0079] Efficient processing and real-time feedback: With the optimized deep neural network and adaptive enhancement strategy, the algorithm can process images quickly and accurately, ensuring that each frame of the image is optimized in real-time processing. It is especially suitable for application scenarios that require real-time feedback such as drone patrol, improving the operation efficiency.

[0080] Robustness and Scene Adaptability: This algorithm not only effectively solves the backlight problem but also can handle other complex lighting changes (such as strong light, shadows, etc.). By optimizing the lighting enhancement curve through multiple iterations, the algorithm has a powerful adaptability and can maintain the image quality in a dynamic lighting environment, ensuring that the target is always clearly visible and centered in the frame.

[0081] Intelligent Image Enhancement: Through a delicately designed lighting enhancement curve and local area enhancement mechanism, this algorithm can restore image details. Especially in the case of target occlusion or complex backgrounds, it can highlight the target and effectively suppress background noise, improving the stability of target recognition and tracking.

[0082] Low Computational Resource Consumption: Based on an optimized network structure and an efficient computing strategy, this algorithm can operate stably in environments with limited computational resources (such as embedded systems or mobile devices). This feature makes the algorithm not only applicable to high-performance computing platforms but also capable of efficient execution on low-power devices, expanding its application scope.

[0083] During the implementation process, the UAV image acquisition system consists of a high-definition camera and sensors mounted on the UAV, which is used to capture image data under different lighting conditions. The original image data obtained by the image acquisition system will be enhanced by this algorithm to ensure the clear presentation of the target in the image and effectively handle complex lighting environments.

[0084] The image acquisition system captures low-light and backlight images through a high-dynamic-range camera, providing image data input. By using an adaptive lighting enhancement high-order curve, it can adjust the brightness and details of the low-light and high-light regions in real time, ensuring that the target in the image is always clearly visible. Especially in backlight or strong light environments, it can effectively restore the occluded details and avoid problems such as overexposure and excessive shadows.

[0085] Lighting Adjustment and Curve Optimization: During the image enhancement process, the algorithm precisely adjusts the brightness of each pixel by iteratively optimizing the adaptive lighting enhancement high-order curve according to different lighting conditions. By introducing an enhancement matrix, the algorithm can apply different intensities of enhancement according to different image regions, avoiding the generation of overexposed or underexposed areas and maintaining the naturalness and details of the image.

[0086] Adaptive Brightness Adjustment of the Target Region: In response to lighting changes in a dynamic environment, the algorithm can adaptively adjust the brightness of the target region to ensure that the target is always clearly visible. By precisely adjusting the brightness and contrast of each region of the image, the algorithm effectively restores the details of the low-light region and improves the overall quality of the image. Especially in the processing of backlight images, the algorithm performs particularly well and can maintain stable image quality in an environment with large lighting changes.

[0087] Real-time Feedback and Control: The system continuously optimizes the image processing process through real-time feedback. During the flight of the drone, the light enhancement curve will be dynamically adjusted according to the real-time image feedback to ensure the stability and continuity of the image. The algorithm can automatically adjust the image enhancement parameters according to the changes in the ambient light of the aircraft, ensuring that the image quality is always in the best state, thereby improving the accuracy of target detection and recognition.

[0088] By applying this algorithm, the drone can continuously and stably capture clear images in environments with backlight, strong light, or dynamic light changes, ensuring that the image quality always remains in the best state under various complex lighting conditions. This algorithm not only effectively improves the accuracy of target recognition and tracking but also can automatically adapt to different lighting environments, reducing the problem of target recognition failure caused by poor image quality. Especially in dynamic environments, the system can adjust the light enhancement parameters in real time, improving the visibility and recognition accuracy of the target, and is suitable for precise monitoring and target positioning tasks, with application potential in multiple fields. For example, the application scenarios include:

[0089] Drone Inspection: This algorithm can be widely applied to image enhancement in drone inspections of infrastructure such as power and communication, ensuring the stability of image quality under backlight or strong light conditions and improving the accuracy of target detection.

[0090] Agricultural Monitoring: In agricultural monitoring, this algorithm can help drones obtain clear data on the growth of crops under different lighting conditions. By enhancing the image quality, it helps farm managers more accurately monitor the health status of crops.

[0091] Intelligent Security: In security monitoring, this algorithm helps drones provide high-quality video streams in complex lighting environments, enhancing the ability to identify potential threats, especially in important applications in fields such as urban and public security.

[0092] In the description of this specification, the descriptions referring to terms such as "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0093] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An image enhancement method applicable to surveillance drones, characterized in that, Including: Estimate the input adaptive illumination enhancement high-order curve through the depth illumination curve estimation network. Through the adaptive illumination enhancement high-order curve, adjust the brightness and details of the low-light and high-light regions in real time to ensure that the targets in the image are always clearly visible, and obtain the enhanced image; The adaptive illumination enhancement high-order curve is: E n (x,y) = E n-1 (x,y) + A n E n-1 (x,y)(1 - E n-1 (x,y)) Among them, (x, y) are the pixel coordinates in the image, and E n (x, y) is the result after the image is enhanced, n represents the number of iterations, and A n is the parameter matrix, whose size is the same as that of the input image; The loss function of the depth illumination curve estimation network is the sum of the reference-free image loss function and the content loss function; the content loss function is used to calculate the pixel-level difference between the generated image and the target image at the corresponding layer, and the result is used as the loss value; The reference-free image loss function is the sum of the spatial consistency loss, the exposure control loss, the color constancy loss, and the illumination smoothness loss; The spatial consistency loss is to maintain the spatial consistency of the image by restricting the difference between adjacent regions of the input image and the enhanced image; The exposure control loss suppresses the exposure degree of the image by controlling the distance between the intensity mean of the local region of the enhanced image and the ideal exposure level; The color constancy loss controls the color deviation between the enhanced image and the input image by establishing the relationship between the three RGB channels; The illumination smoothness loss restricts the mean of the horizontal and vertical gradients in all channels and the number of iterations to maintain the monotonic relationship between adjacent pixels.

2. The image enhancement method for a surveillance drone according to claim 1, characterized in that, The depth illumination curve estimation network estimates the enhancement curve of the input image through a convolutional neural network.

3. An image enhancement method applicable to a surveillance drone according to claim 2, characterized in that, The depth illumination curve estimation network consists of seven symmetrically cascaded convolutional layers. The first six layers are composed of the convolution Conv and the ReLU activation function, which are used to extract the features of the image; Among them, the outputs of the first three convolutional layers are cascaded at the channel level with the outputs of the last three convolutional layers to make full use of the feature information at different levels; the last layer uses the tanh activation function, and finally outputs a set of pixel-level curve parameters describing the corresponding high-order brightness curve.

4. An image enhancement method applicable to a monitoring drone according to claim 3, characterized in that, Each convolutional layer uses a 3×3×32 convolutional kernel.

5. A method for image enhancement applicable to a surveillance drone according to claim 1, characterized in that, The calculation method of the spatial consistency loss is: Among them, K is the number of local regions, Ω(i) represents the four adjacent regions centered on the region, and Y and I are the intensity means of the local regions of the enhanced image and the input image respectively.

6. The image enhancement method for a surveillance drone according to claim 5, wherein, The calculation method of the exposure control loss is: Among them, H is taken as 0.6, and the formula is as follows, where M is the number of local regions, with a size of 16×16, and Y k is the average intensity of the enhanced local region of the image.

7. An image enhancement method applicable to a monitoring drone according to claim 6, characterized in that, The calculation method of the color constancy loss is: Among them, where J p represents the intensity mean of the enhanced image channel p, and (p, q) represents the paired channels of the enhanced image and the input image.

8. An image enhancement method applicable to a surveillance drone according to claim 7, characterized in that, The calculation method of the illumination smoothness loss is: where N is the number of iterations, and represent the gradient calculations in the horizontal and vertical directions, respectively.

9. An image enhancement method for a surveillance drone according to claim 8, characterized in that, The reference-free image loss function is: L total = L spa + L exp + w col L col + w tvA L tvA where w col and w tvA are the weights of the color constancy loss and the illumination smoothness loss, respectively.

10. An image enhancement method for a surveillance drone according to claim 9, characterized in that, The calculation formula of the content loss function is: Among them, represents the feature map extracted from the i-th layer of the VGG16 network, I is the input image, G is the generated image, C i , H i , W i are the number of channels, height, and width of the feature map of the i-th layer in sequence.

Images