Unmanned aerial vehicle shooting pod based on deep learning image enhancement technology

Abstract

The invention discloses an unmanned aerial vehicle shooting pod based on a deep learning image enhancement technology, and belongs to the technical field of unmanned aerial vehicle aerial photography equipment, the unmanned aerial vehicle shooting pod comprises a pod shell, heat dissipation holes are formed in the pod shell, a multispectral camera array is integrated in the pod shell, and the multispectral camera array is electrically connected with an edge calculation unit. An image enhancement model is built in the edge calculation unit, and the edge calculation unit is used for processing an image shot by the multispectral camera array; the unmanned aerial vehicle shooting pod based on the deep learning image enhancement technology provided by the invention is suitable for high-quality image acquisition in severe weather environments such as sand and dust, and through collaborative optimization of a hardware structure and a deep learning algorithm, the problems of image degradation and blurring in the sand and dust environments are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle aerial photography equipment, and particularly relates to an unmanned aerial vehicle shooting pod based on deep learning image enhancement technology. BACKGROUND

[0002] In the field of unmanned aerial vehicle aerial photography, when the unmanned aerial vehicle is in a sand-dust, haze and other harsh weather environment, the traditional shooting equipment faces many problems. On the one hand, the images taken will appear blurred, low contrast, loss of details and other situations, which seriously affects the image quality, leading to the inability to obtain clear and accurate information, greatly limiting the application range of unmanned aerial vehicles in harsh environments. On the other hand, existing physical filtering devices, such as filters, protective glass, etc., can only handle light scattering problems to a certain extent and cannot eliminate the interference of sand particles on imaging from the root, making it difficult to meet the actual application requirements. In addition, conventional image enhancement algorithms, such as histogram equalization, filtering, etc., have poor adaptability in complex sand-dust scenes and are prone to produce artifacts or over-enhancement when processing images, which cannot effectively improve the image quality.

[0003] The existing related technologies also have limitations. For example, the patent document with publication number CN112349456A proposes a unmanned aerial vehicle image defogging method based on multi-spectral fusion, but this method is not optimized for sand-dust environments and cannot effectively solve the image degradation problem in sand-dust weather; the patent document with publication number CN108511593B describes a physical dustproof structure of a unmanned aerial vehicle pod, but this pod does not involve image post-processing technology and cannot improve the image quality from the algorithm level. Therefore, there is an urgent need for a unmanned aerial vehicle shooting pod that can obtain clear images in sand-dust and other harsh weather conditions. SUMMARY

[0004] The purpose of the present application is to provide a unmanned aerial vehicle shooting pod based on deep learning image enhancement technology to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides a unmanned aerial vehicle shooting pod based on deep learning image enhancement technology, comprising a pod shell with a multi-layer dustproof structure, a plurality of heat dissipation holes are arranged on the pod shell, a multi-spectral camera array is integrated inside the pod shell, an edge computing unit is electrically connected to the multi-spectral camera array, and an image enhancement model is built-in in the edge computing unit for processing the images taken by the multi-spectral camera array, and the processing process is as follows:

[0006] S1, the multi-spectral camera array synchronously captures the original images;

[0007] S2, the original images are subjected to hardware-triggered image alignment and denoising preprocessing;

[0008] S3, generating a synthetic dust image through an atmospheric scattering model;

[0009] S4, constructing an image enhancement model, using a transfer learning strategy, training the image enhancement model with the synthetic dust image, and fine-tuning the image enhancement model with real dust images;

[0010] S5, inputting the preprocessed image into the image enhancement model for processing;

[0011] S6, fusing the physical model and the statistical characteristics to adaptively optimize the image output by the image enhancement model, and outputting a high-quality and clear image; the atmospheric scattering model directly takes the output of the image enhancement model as input to recalculate the transmittance and clear image, and forcibly corrects the "physical deviation" of color / brightness, for example, if the sky area output by the model is yellow (dust residue), the inversion process will recalibrate the blue channel intensity through A to make the color more realistic. The statistical characteristics optimization takes the image after physical inversion as input to calculate the local contrast and dynamically determine the sharpening intensity, for example, when the contrast of the vegetation area is low, the sharpening weight w is increased to strengthen the leaf texture; when the contrast of the sky area is high, w is reduced to avoid pseudo-texture.

[0012] Preferably, the pod shell is composed of a dust cover, a detachable dust filter screen and a waterproof breathable film, which reduces the entry of dust particles into the multispectral camera array; the multispectral camera array includes an RGB camera and a near-infrared camera, which synchronously collects images of different wavebands to provide multi-dimensional data for subsequent processing.

[0013] Preferably, the hardware-triggered image alignment of step S2 generates a registered image I aligned based on an affine transformation matrix.

[0014]

[0015] wherein N is the number of matched feature points; is a pair of matched feature points in adjacent frames of images;

[0016] Based on the image I aligned , denoising is performed by combining bilateral filtering and non-local mean filtering (NLM):

[0017] I denoised = a·Bilateral(I aligned ) + (1-a)·NLM(I aligned );

[0018] wherein a is an adaptive weight that is dynamically adjusted according to the local noise intensity.

[0019] Preferably, step S3 specifically includes:

[0020] S31, selecting a clear image from a public data set or a collected dust-free image;

[0021] S32, setting the parameter scattering coefficient β, scene depth d(x) and atmospheric light value A to simulate different dust scenes;

[0022] S33, substituting β and d(x) into the formula t(x) = e -β·d(x) Calculate the transmittance pixel by pixel;

[0023] S34, substitute the clear image, t(x) and A into the formula of the atmospheric scattering model, and calculate the dust degradation image pixel by pixel:

[0024] I(x) = J(x) t(x) + A (1-t(x));

[0025] Wherein, J(x) is a clear image.

[0026] Preferably, the image enhancement model in step S4 adopts a DustEnhanceNet architecture model, which includes an encoder, a channel attention interaction module, a multi-scale convolution module and a decoder, specifically:

[0027] The encoder uses MobileNetv2 to extract three scale feature maps, namely high, medium and low scale feature maps;

[0028] The channel attention interaction module fuses multi-scale feature maps;

[0029] The multi-scale convolution module extracts and enhances the feature map details of the shallowest layer, and uses a multi-scale loss function for supervised learning to improve the image clarity;

[0030] The decoder fuses shallow geometric features and deep semantic features through transposed convolution and skip connection, and outputs the final enhanced image.

[0031] Preferably, the channel attention interaction module is specifically:

[0032] Feature fusion, connect high, medium and low scale feature maps:

[0033] x' = Cat(C 1×1 (x high ),C 3×3 (x mid ),C 3×3 (x low ));

[0034] Channel attention calculation:

[0035] ca' = Sigmoid(C 1×1 (GAP(x′)))

[0036] Feature weight interaction:

[0037] x′ high =x high +ca′·C 1×1 (x high );

[0038] x′ mid =x mid +ca′·C 1×1 (x mid );

[0039] Cross-scale fusion:

[0040] x mci =C 3×3 (Cat(x′ high ,x′ mid ,x low ));

[0041] Wherein, x high , x mid , x low are high, medium and low scale feature maps respectively.

[0042] Preferably, the multi-scale convolution module specifically comprises:

[0043] Parallel branch feature extraction:

[0044] x1=C 1×1 (xmci);

[0045] x3=DWC 3×3 (C 3×1 (C 1×1 (x mci )));

[0046] x5=DWC 5×5 (C 5×1 (C 1×1 (x mci )));

[0047] x7=DWC 7×7 (C 7×1 (C 1×1 (x mci )));

[0048] Feature fusion and residual connection:

[0049] x msc =C 1×1 (Cat(x1,x3,x5,x7));

[0050] x out=C 3×3 (x msc )+x mci .

[0051] Preferably, the multi-scale loss function is as follows:

[0052] L=αL Pixel +σL SSIM +γL Grad +δL Perceptual ;

[0053] Among them, L Pixel For L1 loss, L SSIM For structural similarity loss, L Grad For gradient loss, L Perceptual For the perceived loss, α, σ, γ, and δ are all weighting coefficients.

[0054] Preferably, step S6 optimizes the output of the image enhancement model based on atmospheric scattering model inversion and statistical characteristics, specifically as follows:

[0055] Estimated transmittance:

[0056] Restore clear images:

[0057] Calculate local contrast:

[0058] Dynamically adjust sharpening intensity:

[0059] in, The image output by the image enhancement model; for The minimum value in the three channels; ∈ represents the numerical stability term; t0 is the lower limit of transmittance; σ(x) is the standard deviation of a local region of the image (reflecting pixel value fluctuations, i.e., contrast levels); μ(x) is the mean of a local region of the image (reflecting overall brightness); I final For a clear final output image; τ is the Laplacian operator for the image; τ is the baseline value for sharpening intensity; w is a dynamic weight determined by local contrast (w→1 in low-contrast areas to enhance details; w→0 in high-contrast areas to maintain a natural look), and k is the scaling factor of the Sigmoid function, which controls the sensitivity of contrast to the weight w (the larger k is, the closer w in low-contrast areas is to 1, and the stronger the sharpening).

[0060] Therefore, the UAV imaging pod based on deep learning image enhancement technology described above has the following beneficial effects:

[0061] (1) Effectively solve the image degradation problem under sand-dust weather;

[0062] (2) Through transfer learning + physical inversion, the model simultaneously learns the "physical law of synthetic data" and the "scene specificity of real data";

[0063] (3) DustEnhanceNet is improved based on U-Net, which can flexibly replace the backbone network (such as replacing EfficientNet), expand multispectral channels, and adapt to more scenes;

[0064] (4) Through deep learning + physical model + hardware cooperation, the breakthrough of taking good images and clear images in harsh environments is realized.

[0065] The technical solutions of the present application will be further described in detail below through the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 It is a structure schematic diagram of the unmanned aerial vehicle shooting pod based on the deep learning image enhancement technology of the present application;

[0067] Figure 2 It is an image processing overall flowchart of the unmanned aerial vehicle shooting pod based on the deep learning image enhancement technology of the present application;

[0068] Figure 3 It is a flowchart of the synthetic sand-dust image of the embodiment of the present application. DETAILED DESCRIPTION

[0069] The following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0070] Please refer to Figure 1 , an unmanned aerial vehicle shooting pod based on deep learning image enhancement technology, comprising a pod shell with a multi-layer dustproof structure, which is made of carbon fiber composite material, which reduces the weight by 30% compared with aluminum alloy while ensuring the structural strength. The pod shell is provided with a heat dissipation hole, and the pod shell is integrated with a multispectral camera array inside. The pod shell is composed of a dustproof cover, a detachable dust filter screen and a waterproof breathable film, which reduces the entry of dust particles into the multispectral camera array; the multispectral camera array includes an RGB camera and a near-infrared camera, which synchronously collects images of different wavebands to provide multi-dimensional data for subsequent processing. The multispectral camera array is electrically connected with an edge computing unit, and the edge computing unit is built-in with an image enhancement model, which is used to process the images taken by the multispectral camera array, such as Figures 2-3The processing process is as follows:

[0071] S1, a multispectral camera array synchronously captures original images;

[0072] S2, hardware triggered image alignment and denoising preprocessing are performed on the original images; the hardware triggered image alignment generates a registered image I aligned based on an affine transformation matrix, and the affine transformation matrix is:

[0073]

[0074] wherein N is the number of matched feature points; is a matched feature point pair in adjacent frame images;

[0075] Based on the image I aligned , combined with bilateral filtering and non-local mean filtering (NLM), denoising processing is performed:

[0076] I denoised = a·Bilateral(I aligned ) + (1-a)·NLM(I aligned );

[0077] wherein a is an adaptive weight, which is dynamically adjusted according to the local noise intensity.

[0078] S3, a synthetic dust image is generated through an atmospheric scattering model, specifically including:

[0079] S31, a clear image is selected from a public data set or collected dust-free image;

[0080] S32, parameters scattering coefficient β, scene depth d(x) and atmospheric light value A are set to simulate different dust scenes;

[0081] S33, β and d(x) are substituted into the formula t(x) = e -β·d(x) to calculate the transmittance pixel by pixel;

[0082] S34, the clear image, t(x) and A are substituted into the formula of the atmospheric scattering model to calculate the dust degradation image pixel by pixel:

[0083] I(x) = J(x)·t(x) + A·(1-t(x));

[0084] wherein J(x) is the clear image.

[0085] S4, construct an image enhancement model, adopt a transfer learning strategy, train the image enhancement model using synthetic dust images, and fine-tune the image enhancement model using real dust images; the image enhancement model adopts a DustEnhanceNet architecture model, including an encoder, a channel attention interaction module, a multi-scale convolution module, and a decoder, specifically:

[0086] The encoder adopts MobileNetv2 to extract three scale feature maps, namely high, medium, and low scale feature maps.

[0087] The channel attention interaction module fuses multi-scale feature maps, specifically:

[0088] Feature fusion, connecting high, medium, and low scale feature maps:

[0089] x' = Cat(C 1×1 (x high ), C 3×3 (x mid ), C 3×3 (x low ));

[0090] Channel attention calculation:

[0091] ca' = Sigmoid(C 1×1 (GAP(x′)));

[0092] Feature weight interaction:

[0093] x' high = x high + ca' · C 1×1 (x high );

[0094] x' mid = x mid + ca' · C 1×1 (x mid );

[0095] Cross-scale fusion:

[0096] x mci = C 3×3 (Cat(x' high ,x' mid ,x low ));

[0097] Wherein, x high , x mid , x low are high, medium, and low scale feature maps.

[0098] Multi-scale convolution module, extracts and enhances the feature map details of the shallowest layer, and uses a multi-scale loss function for supervised learning to improve image clarity, which specifically includes:

[0099] Parallel branch feature extraction:

[0100] x1=C 1×1 (x mci );

[0101] x3=DWC 3×3 (C 3×1 (C 1×1 (x mci )));

[0102] x5=DWC 5×5 (C 5×1 (C 1×1 (x mci )));

[0103] x7=DWC 7×7 (C 7×1 (C 1×1 (x mci )));

[0104] Feature fusion and residual connection:

[0105] x msc =C 1×1 (Cat(x1,x3,x5,x7));

[0106] x out =C 3×3 (x msc )+x mci .

[0107] The multi-scale loss function is specifically:

[0108] L=αL Pixel +σL SSIM +γL Grad +δL Perceptual ;

[0109] Wherein, L Pixel is the L1 loss, L SSIM is the structural similarity loss, L Grad is the gradient loss, L Perceptual is the perception loss, and α, σ, γ, δ are weight coefficients. Among them:

[0110] L1 loss: Optimize color restoration, M is the number of pixels.

[0111] Structural similarity loss L SSIM : Improve clarity.

[0112] Gradient loss L Grad : Preserve details This is the Sobel operator.

[0113] Perceived loss L Perceptual : To ensure semantic consistency, φ is the feature extraction layer of the pre-trained VGG network.

[0114] The decoder, through transposed convolution and skip connections, fuses shallow geometric features with deep semantic features to output the final enhanced image.

[0115] S5. Input the preprocessed image into the image enhancement model for further processing;

[0116] S6. The image output from the image enhancement model is adaptively optimized by fusing physical models and statistical features to produce a high-quality, clear image. Atmospheric scattering model inversion directly uses the output of the image enhancement model as input, recalculating transmittance and image clarity, and forcibly correcting "physical deviations" in color / brightness. For example, if the sky area output by the model is yellowish (due to residual dust), the inversion process will recalibrate the blue channel intensity through A to make the colors more realistic. Statistical feature optimization uses the physically inverted image as input, calculates its local contrast, and dynamically determines the sharpening intensity. For example, when the contrast in the vegetation area is low, the sharpening weight w is increased to enhance leaf texture; when the contrast in the sky area is high, w is decreased to avoid false textures. Specifically:

[0117] Estimated transmittance:

[0118] Restore clear images:

[0119] Calculate local contrast:

[0120] Dynamically adjust sharpening intensity:

[0121] in, The image output by the image enhancement model; for The minimum value in the three channels; ∈ represents the numerical stability term; t0 is the lower limit of transmittance; σ(x) is the standard deviation of a local region of the image; μ(x) is the mean of a local region of the image; I final For a clear final output image; Let be the Laplacian operator for the image; τ be the baseline value for sharpening intensity; and w be the dynamic weights. k is the scaling factor of the Sigmoid function.

[0122] This embodiment selects a certain multi-rotor unmanned aerial vehicle, carries a customized shooting pod, integrates a multispectral camera (resolution 4K, frame rate 30 fps), an FPGA preprocessing module, and an NVIDIA Jetson AGX Orin edge computing unit (for deploying the DustEnhanceNet model); environmental simulation equipment: sand-dust environment simulation cabin (controllable sand-dust concentration and distribution), high-low temperature test box (temperature can be set to -20℃-60℃), and dustproof test device (detection according to IP67 standard).

[0123] The classical defogging algorithm (such as the dark channel prior algorithm) and the traditional image enhancement algorithm (such as histogram equalization) are selected as the comparison, and the algorithms of this embodiment are all based on the PyTorch framework to build and train the model, and are deployed to the edge computing unit through TensorRT acceleration to realize real-time inference. The experimental process is as follows:

[0124] (I) Data set construction

[0125] Synthetic sand-dust images: use the atmospheric scattering model I(x) = J(x) t(x) + A (1-t(x)), t(x) = e -β·d(x) Generate a synthetic data set. Set the parameters: randomly select 12 RGB combinations for the atmospheric light value A (simulate different sand-dust tones), take 0.1-0.5 for the scattering coefficient β (simulate different sand-dust concentrations), and generate 100,000+ “clear image-sand-dust image” sample pairs in combination with the NYU Depth V2 depth data set.

[0126] Real sand-dust images: collect 1000+ real sand-dust images through the unmanned aerial vehicle pod in desert and construction dust scenes to construct a real data set for model fine-tuning and testing.

[0127] (II) Model training and optimization

[0128] Pre-training: train the DustEnhanceNet model on the synthetic data set, with 50 training rounds, a batch size (BatchSize) of 8, an Adam optimizer, an initial learning rate of 0.001, and a multi-scale loss function L = 0.5L Pixel + 0.3L SSIM + 0.1L Grad + 0.1L Perceptual , so that the model learns the physical law of sand-dust degradation.

[0129] Fine-tuning: fine-tune the model on the real sand-dust data set, with 10 training rounds, a batch size of 8, and a learning rate adjustment of 0.0001, to adapt to the real environment noise and sand-dust distribution characteristics and improve the robustness of the model.

[0130] (III) Image enhancement effect experiment

[0131] Test dataset: 200 images are selected from real dust dataset, covering different dust concentrations (light, medium, heavy), scenes (desert, urban dust).

[0132] Experimental procedure:

[0133] The test images are processed using the traditional method (dark channel prior + histogram equalization) and DustEnhanceNet, respectively.

[0134] Based on PSNR and SSIM evaluation, PSNR is calculated according to (MAX I , MSE is the mean square error), SSIM is calculated according to (c1=(0.01L) 2 , c2=(0.03L) 2 , L=255).

[0135] Experimental results:

[0136] The average PSNR of the traditional method is 20.5 dB, and the average PSNR of DustEnhanceNet is 30.8 dB, with a clarity improvement of about 50.2%.

[0137] The average SSIM of the traditional method is 0.72, and the average SSIM of DustEnhanceNet is 0.94, with a detail retention improvement of about 30.6%.

[0138] (IV) Environmental adaptability experiment

[0139] Wide temperature experiment:

[0140] The UAV with the pod is placed in a high and low temperature test box, and temperature points of -20℃, 0℃, 25℃, 40℃, 60℃ are set. After stabilizing for 30 minutes at each temperature point, images are collected and the pod working state is tested.

[0141] Test content: Check if the camera and processing unit are working normally, test if the image enhancement function is stable, and record the end-to-end processing delay.

[0142] Experimental results: At each temperature point, the pod can work normally, the image enhancement effect is stable, the end-to-end processing delay is <50ms, which meets the wide temperature working requirements.

[0143] Dustproof experiment:

[0144] According to the IP67 dustproof standard, the pod is placed in a dustproof test device to simulate a sand dust environment (dust concentration 200mg / m 3 ) for 24 hours of continuous testing.

[0145] Post-test inspection: Check if dust enters the pod, whether the camera lens, processing unit and other key components are normal, and test the image acquisition and enhancement functions.

[0146] Experimental results: There is no obvious dust entering the pod, the key components are working normally, the image acquisition and enhancement functions are stable, and the IP67 dustproof level is reached.

[0147] (Five) Real-time experiment

[0148] Test scene: In an open outdoor site, the unmanned aerial vehicle flies at a speed of 5m / s, continuously collects video stream (frame rate 30fps), and tests the end-to-end processing delay (from image acquisition to enhanced image output time).

[0149] Test method: Through the timestamp module built-in pod, record the image acquisition time and enhanced output time, calculate the delay, and count the average delay of 1000 frames of images.

[0150] Experimental results: The average end-to-end processing delay is 42ms, less than 50ms, which meets the real-time aerial photography requirements.

[0151] Four, experimental conclusion

[0152] This embodiment trains the DustEnhanceNet model by constructing synthetic and real data sets, and verifies it through multi-dimensional experiments:

[0153] On image enhancement effect, compared with traditional scheme, the clarity (PSNR) is improved by more than 40%, the detail retention (SSIM) is improved by more than 30%, and the aerial image quality in dust environment is effectively enhanced; On environmental adaptability, it supports wide temperature work of-20℃-60℃, reaches IP67 dustproof level, and adapts to harsh aerial photography scene; On real-time, the end-to-end processing delay is less than 50ms, which meets the real-time aerial photography requirements of unmanned aerial vehicle, and verifies the feasibility and advantages of the scheme.

[0154] Therefore, the unmanned aerial vehicle shooting pod based on the deep learning image enhancement technology is adopted, which is suitable for high-quality image acquisition in dust and other harsh weather environments, and through the cooperative optimization of hardware structure and deep learning algorithm, the problem of image degradation in dust environment is solved.

[0155] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalent ones, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A drone imaging pod based on deep learning image enhancement technology, characterized in that, The system includes a pod shell with ventilation holes. A multispectral camera array is integrated inside the pod shell. The multispectral camera array is electrically connected to an edge computing unit. The edge computing unit has a built-in image enhancement model for processing images captured by the multispectral camera array. The processing procedure is as follows: S1. Multispectral camera array synchronously captures raw images; S2. Perform hardware-triggered image alignment and denoising preprocessing on the original image; S3. Generate synthetic dust images using an atmospheric scattering model; S4. Construct an image enhancement model, adopt a transfer learning strategy, train the image enhancement model using synthetic sand and dust images, and then fine-tune the image enhancement model using real sand and dust images; S5. Input the preprocessed image into the image enhancement model for further processing; S6. The physical model and statistical features are integrated to adaptively optimize the image output by the image enhancement model, resulting in a high-quality and clear image.

2. The UAV shooting pod based on deep learning image enhancement technology according to claim 1, characterized in that: The pod's outer shell consists of a dust cover, a removable sand and dust filter, and a waterproof and breathable membrane; the multispectral camera array includes an RGB camera and a near-infrared camera.

3. The UAV shooting pod based on deep learning image enhancement technology according to claim 1, characterized in that, Step S2, hardware-triggered image alignment, generates the registered image I based on the affine transformation matrix. aligned The affine transformation matrix is: Where N is the number of matching feature points; These are pairs of feature points matched in adjacent frames of the image. Based on image I aligned The noise reduction process combines bilateral filtering and nonlocal mean filtering (NLM). I denoised =α·Bilateral(I aligned )+(1-α)·NLM(I aligned ); Where α is the adaptive weight.

4. The UAV shooting pod based on deep learning image enhancement technology according to claim 1, characterized in that, Step S3 specifically includes: S31. Select clear images from publicly available datasets or collected dust-free images; S32. Set the parameters scattering coefficient β, scene depth d(x) and atmospheric light value A to simulate different sandstorm scenarios; S33. Substitute β and d(x) into the formula t(x) = e -β·d(x) Calculate transmittance pixel by pixel; S34. Substitute the clear image, t(x), and A into the formula of the atmospheric scattering model to calculate the dust degradation image pixel by pixel: I(x)=J(x)·t(x)+A·(1-t(x)); Where J(x) represents the sharp image.

5. The UAV shooting pod based on deep learning image enhancement technology according to claim 1, characterized in that, The image enhancement model in step S4 adopts the DustEnhanceNet architecture, which includes an encoder, a channel attention interaction module, a multi-scale convolution module, and a decoder, specifically: The encoder uses MobileNetv2 to extract feature maps at three scales: high, medium, and low scales. Channel attention interaction module, which integrates multi-scale feature maps; The multi-scale convolution module extracts and enhances the details of the shallowest feature map, and uses a multi-scale loss function for supervised learning to improve image clarity. The decoder, through transposed convolution and skip connections, fuses shallow geometric features with deep semantic features to output the final enhanced image.

6. The UAV shooting pod based on deep learning image enhancement technology according to claim 5, characterized in that, The channel attention interaction module is as follows: Feature fusion, connecting high-, medium-, and low-scale feature maps: x'=Cat(C 1×1 (x high ),C 3×3 (x mid ),C 3×3 (x low )); Channel attention calculation: ca'=Sigmoid(C 1×1 (GAP(x'))); Feature weight interaction: x' high =x high +ca'·C 1×1 (x high ); x' mid =x mid +ca'·C 1×1 (x mid ); Cross-scale fusion: x mci =C 3×3 (Cat(x' high ,x' mid ,x low )); Where, x high x mid x low These are high-, medium-, and low-scale feature maps, respectively.

7. The UAV shooting pod based on deep learning image enhancement technology according to claim 6, characterized in that, The multi-scale convolution module specifically includes: Parallel branch feature extraction: x1=C 1×1 (x mci ); x3=DWC 3×3 (C 3×1 (C 1×1 (x mci ))); x5=DWC 5×5 (C 5×1 (C 1×1 (x mci ))); x7=DWC 7×7 (C 7×1 (C 1×1 (x mci ))); Feature fusion and residual connectivity: x msc =C 1×1 (Cat(x1,x3,x5,x7)); x out =C 3×3 (x msc )+x mci 。 8. The UAV shooting pod based on deep learning image enhancement technology according to claim 7, characterized in that, The multi-scale loss function is as follows: L=αL Pixel +σL SSIM +γL Grad +δL Perceptual ; Among them, L Pixel For L1 loss, L SSIM For structural similarity loss, L Grad For gradient loss, L Perceptual For the perceived loss, α, σ, γ, and δ are all weighting coefficients.

9. A drone photography pod based on deep learning image enhancement technology according to claim 8, characterized in that, Step S6 optimizes the output of the image enhancement model based on atmospheric scattering model inversion and statistical features, specifically as follows: Estimated transmittance: Restore clear images: Calculate local contrast: Dynamically adjust sharpening intensity: in, The image output by the image enhancement model; for The minimum value in the three channels; ∈ represents the numerical stability term; t0 is the lower limit of transmittance; σ(x) is the standard deviation of a local region of the image; μ(x) is the mean of a local region of the image; I final For a clear final output image; Let be the Laplacian operator for the image; τ be the baseline value for sharpening intensity; and w be the dynamic weights. k is the scaling factor of the Sigmoid function.

Images