An image super-resolution reconstruction method

Abstract

The application discloses an image super-resolution reconstruction method, which comprises the following steps: acquiring original image point cloud data, generating a sparse depth projection graph containing depth values and reflection intensity; extracting visual environment features of the original image and physical reflection features of the point cloud data, constructing a cross-modal conflict arbitration mechanism, and obtaining a to-be-reconstructed image; generating a computing power allocation mask according to the depth values in the sparse depth projection graph and a preset safe attention distance threshold value, and marking a region with a depth value greater than the distance threshold value as a high-attention reconstruction region; constructing a super-resolution network model containing a depth index dynamic sparse convolution layer; inputting the to-be-reconstructed image into the super-resolution network model, controlling an inference path by using the computing power allocation mask, calling an expansion rate matched with the depth of the high-attention reconstruction region for feature extraction only, and outputting a final image. The application can solve the problems that the prior art is difficult to balance inference speed and reconstruction accuracy, has high invalid power consumption, and has a high environment perception misjudgment rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an image super-resolution reconstruction method. BACKGROUND

[0002] In the automatic driving and auxiliary driving system, the perception accuracy of long-distance targets (such as traffic signs 100 meters away) is directly related to driving safety. Although image super-resolution technology can restore details, it still faces serious bottlenecks in actual vehicle engineering landing:

[0003] 1. Existing technologies are difficult to balance inference speed and reconstruction accuracy: existing high-performance SR algorithms (such as RCAN, EDSR) have extremely high computational complexity, and when running on a vehicle edge computing platform with limited computing power (such as early Orin-Nano or FPGA chips), it will cause serious inference delay, which cannot meet the real-time requirements. Although there are also some lightweight super-resolution models (such as FSRCNN), the reconstruction details of long-distance small targets (such as road sign text 100 meters away) are severely lost, which cannot meet the stringent requirements of automatic driving for perception accuracy.

[0004] 2. Invalid power consumption is too high: traditional convolutional networks use force on all pixels in the full image, and driving scenes contain a large amount of background such as sky and nearby road surface that do not need to be concerned. High-power reconstruction of these areas will generate a large amount of invalid power consumption, aggravate chip heating, and even trigger the system's frequency reduction protection.

[0005] 3. High environment perception misjudgment rate: in complex white walls, snow-covered ground, or high-reflectivity road surfaces, and other special scenes, a single visual perception algorithm is prone to environmental misjudgment (for example, mistaking a white wall for heavy fog) due to feature confusion. This misjudgment will cause the system to incorrectly activate the de-fog enhancement process, not only wasting computing resources, but also introducing visual artifacts that do not exist in the image, interfering with subsequent perception tasks. In addition, existing multi-modal fusion schemes mostly use "feature-level cascading", which not only increases the amount of calculation, but also cannot effectively arbitrate when radar and visual features are inconsistent (such as in rainy or strong light), and may instead introduce noise. SUMMARY

[0006] In view of this, the present application provides an image super-resolution reconstruction method to solve the problems of existing technologies being difficult to balance inference speed and reconstruction accuracy, invalid power consumption being too high, and high environment perception misjudgment rate.

[0007] An image super-resolution reconstruction method, comprising:

[0008] Step S1: Obtain the original image captured by the vehicle-mounted camera and the point cloud data of the LiDAR, project the point cloud data onto the pixel coordinate system of the original image, and generate a sparse depth projection map containing depth values ​​and reflection intensity.

[0009] Step S2: Extract the visual environment features of the original image and the physical reflection features of the point cloud data, and construct a cross-modal conflict arbitration mechanism. If the visual environment features indicate the presence of environmental interference, but the physical reflection features indicate the absence of environmental interference, the current scene is determined to be a false fog scene. No dehazing preprocessing is performed, and the original image is directly output as the image to be reconstructed. If both the visual environment features and the physical reflection features indicate the presence of environmental interference, the current scene is determined to be a real fog scene. Dehazing preprocessing is performed on the original image, and the generated dehazing enhanced image is used as the image to be reconstructed.

[0010] Step S3: Based on the depth value in the sparse depth projection map and the preset safety concern distance threshold, generate a computing power allocation mask, mark the region with a depth value greater than the distance threshold as a high concern reconstruction region, and perform a morphological expansion operation on the high concern reconstruction region to make the high concern reconstruction region form a physically continuous spatial structure.

[0011] Step S4: Construct a super-resolution network model containing a depth-indexed dynamic sparse convolutional layer. In this model, establish a step mapping relationship between depth range and dilation rate. From the preset depth-dilation rate lookup table, use the depth value of the sparse depth projection map as an index to retrieve the corresponding dilation rate. Dynamically load the convolutional kernel weights that match the dilation rate from the shared weight pool to achieve adaptive feature sampling based on physical depth.

[0012] Step S5: Input the image to be reconstructed into the super-resolution network model, use computing power to allocate masks to control the inference path, only call the dilation rate matching its depth for feature extraction in the high-interest reconstruction area, and perform downsampling or interpolation processing on the non-high-interest reconstruction area, and then output the final image.

[0013] The image super-resolution reconstruction method provided by the present invention has the following beneficial effects:

[0014] (1) To address the issues of blurred details in existing lightweight models when dealing with small targets at long distances, and excessive inference latency in high-performance models, this invention proposes a dynamic sparse convolutional layer based on depth indexing. This layer establishes a mapping relationship between depth intervals and dilation rates, and can adaptively adjust the receptive field size according to the physical distance of the target: standard convolution can be used to preserve texture at close range, while convolution with a large dilation rate can be used to enhance contours at long range. Combined with a dynamic dilation rate decoupling training strategy, this invention effectively improves the super-resolution reconstruction clarity of key targets such as traffic signs at a distance of 100 meters without significantly increasing the number of model parameters. Therefore, this invention balances inference speed and reconstruction accuracy.

[0015] (2) To address the problem that traditional convolutional networks perform equal calculations on all pixels of the image, resulting in a large amount of ineffective power consumption in background areas such as the sky and nearby road surfaces, this invention utilizes the depth information of LiDAR to generate a computing power allocation mask. By introducing a spatial sparsity mechanism during network inference, downsampling or interpolation is performed on non-high-interest reconstruction areas, while high-resolution reconstruction is performed only on high-interest reconstruction areas containing distant key targets. This physically guided computing strategy significantly reduces the floating-point operation volume and chip power consumption of the model, achieving efficient allocation of onboard computing resources.

[0016] (3) To address the problem that existing single-vision solutions are prone to misjudging "false fog" in scenarios such as white walls and highly reflective surfaces, leading to incorrect activation of defogging processing and the introduction of noise, this invention constructs a cross-modal conflict arbitration mechanism based on physical features. This mechanism utilizes the physical reflection characteristics of point cloud data as an objective physical criterion to logically verify visual environment features. This mechanism effectively distinguishes between real aerosol scattering and high surface reflectivity, activating enhancement processing only when both visual environment features and physical reflection features indicate the presence of environmental interference. This avoids the waste of computational power and image distortion risks caused by algorithmic misjudgment at the source, reducing the environmental perception misjudgment rate. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the image super-resolution reconstruction method provided in an embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0019] Please see Figure 1 The image super-resolution reconstruction method provided by the present invention includes steps S1 to S5:

[0020] Step S1: Obtain the original image captured by the vehicle-mounted camera and the point cloud data of the LiDAR, project the point cloud data onto the pixel coordinate system of the original image, and generate a sparse depth projection map containing depth values ​​and reflection intensity.

[0021] The onboard processor first uses a vehicle kinematics model to perform motion compensation on the low-frequency point cloud data, aligning its timestamps with the image frames of the original image. Then, based on a pre-calibrated camera-radar extrinsic matrix and camera intrinsic parameters, the 3D point cloud data is projected onto a 2D image pixel coordinate system.

[0022] Specifically, the process of projecting point cloud data onto the pixel coordinate system of the original image satisfies the following equation:

[0023]

[0024] in, , , The coordinates are three-dimensional coordinates in the lidar coordinate system. , These are the pixel coordinates of the point cloud data projected onto the original image. Depth in camera coordinate system This is the camera's internal parameter matrix (including focal length and optical center coordinates). Let be the rotation matrix of the lidar coordinate system relative to the camera coordinate system. Let be the translation vector of the lidar relative to the camera.

[0025] In the specific calculation, the point cloud data is first transformed into the camera coordinate system. Then, perspective projection is performed using camera intrinsic parameters, and the result is divided by the depth value (i.e., homogeneous coordinate normalization) to calculate the corresponding image pixel coordinates. For points whose projection results fall within the image resolution range, they are extracted. The value is used as the depth value And retain the corresponding reflection intensity Generate a sparse depth projection map; for areas without point cloud coverage, set the depth value to zero.

[0026] Step S2: Extract the visual environment features of the original image and the physical reflection features of the point cloud data, and construct a cross-modal conflict arbitration mechanism. If the visual environment features indicate the presence of environmental interference, but the physical reflection features indicate the absence of environmental interference, the current scene is determined to be a false fog scene. No dehazing preprocessing is performed, and the original image is directly output as the image to be reconstructed. If both the visual environment features and the physical reflection features indicate the presence of environmental interference, the current scene is determined to be a real fog scene. Dehazing preprocessing is performed on the original image, and the generated dehazing enhanced image is used as the image to be reconstructed.

[0027] Specifically, step S2 includes:

[0028] The normalized dark channel average intensity of the original image is calculated to obtain the visual fog index. Visual fog index As a feature of the visual environment.

[0029] The reflection intensity and distance information of point cloud data are extracted, and the spatial attenuation slope of the reflection intensity is fitted by linear regression. Specifically, first, we need to analyze the reflection intensity of the point cloud data. With depth value For linear regression, the expression is: ,in, The intercept term characterizes the combined response of the target object's base reflection intensity and radar transmit power. It is then fitted using the least squares method or random sample-consensus algorithm to obtain the spatial attenuation slope. , This is used to characterize the attenuation effect of aerosol concentration on the light beam, and then based on the spatial attenuation slope. Calculate the radar fog perception index Radar fog perception index Calculated using the following formula:

[0030]

[0031] in, The preset sensitivity coefficient is preferably 10.0-15.0; This is the preset baseline attenuation threshold.

[0032] Radar fog perception index As a physical reflection characteristic.

[0033] Arbitration Difference Values ​​for Constructing Cross-Modal Conflict Arbitration Mechanisms , .

[0034] In this embodiment, the visual fog index The necessity of this preliminary judgment condition lies in the fact that while lidar has the advantage of physical ranging, its backscattering response to low-concentration aerosols (such as fog or light haze) is weak, and the spatial resolution of point clouds is much lower than that of images, making it difficult to accurately perceive changes in visibility across the entire image. Relying solely on lidar data can easily lead to missed detections of foggy scenes, resulting in images not being enhanced when they should be. Therefore, this embodiment employs a closed-loop logic of first capturing suspected fog features using visual algorithms and then using lidar data for physical verification, ensuring that the system neither overlooks fog nor mistakenly detects white walls.

[0035] Scenario A: When the visual fog index Greater than or equal to a preset visual threshold (preferably 0.6 in this embodiment), and When the collision threshold is greater than or equal to the threshold, it indicates that although the visual features resemble fog, the radar wave transmission is not attenuated by aerosol scattering (typically corresponding to white buildings or highly reflective surfaces). The current scene is determined to be a false fog scene. Based on the physical reliability of the LiDAR, a control command is generated to skip the defogging preprocessing (i.e., execute the bypass strategy). This means that defogging preprocessing is not performed, and the original image is directly output as the image to be reconstructed and transmitted to subsequent reconstruction processes. This strategy avoids the block artifacts and color distortion introduced by forcibly defogging non-foggy images at the source, ensuring the input fidelity of subsequent super-resolution reconstruction.

[0036] Scenario B: When the visual fog index Greater than or equal to a preset visual threshold, and When the value is less than the conflict threshold, it indicates that the visually perceived fog characteristics and the intensity attenuation characteristics of radar waves have physical consistency. The current scene is determined to be a real fog scene. Dehazing preprocessing is performed on the original image. Through optical restoration, the generated dehazing enhanced image is used as the image to be reconstructed and transmitted to the subsequent reconstruction process to eliminate the occlusion of image details by environmental scattering.

[0037] Scenario C: When the visual fog index When the image is less than the preset visual threshold, regardless of the radar data, the current scene is determined to be a scene without visual interference. No dehazing preprocessing is performed, and the original image is directly output as the image to be reconstructed, thus avoiding unnecessary signal processing on the clear image.

[0038] Step S3: Based on the depth value in the sparse depth projection map and the preset safety concern distance threshold, a computing power allocation mask is generated. Regions with depth values ​​greater than the distance threshold are marked as high concern reconstruction regions. Morphological expansion operation is performed on the high concern reconstruction regions to form a physically continuous spatial structure.

[0039] In this embodiment, considering the vehicle braking distance, the distance threshold is... Set to 20 meters (this threshold can be dynamically adjusted according to vehicle speed).

[0040] For depth value The projection point region, in the computing power allocation mask The region marked as "1" is designated as the high-interest reconstruction region. A morphological dilation operation is performed on the high-interest reconstruction region, specifically a 3x3 morphological dilation operation, to form a physically continuous spatial structure in the high-interest reconstruction region, preventing the structural breakage of the target object due to sparse sampling.

[0041] For depth value In areas with no data, the computing power allocation mask is used. The region marked as "0" is considered a non-high-interest reconstruction region. During network inference, convolutional feature extraction is skipped for non-high-interest reconstruction regions, and zero values ​​are directly output or bilinear interpolation is performed, retaining only the super-resolution reconstruction results for high-interest reconstruction regions. This is because near-field targets (e.g., within 20 meters) already occupy sufficient pixel area in the original image captured by the vehicle-mounted camera, and their physical resolution usually meets the feature extraction requirements of the perception algorithm, eliminating the need for additional super-resolution reconstruction. Therefore, performing interpolation on non-high-interest reconstruction regions can both preserve the original visual information to ensure the safety of near-field perception and avoid high computational consumption in unnecessary areas.

[0042] Step S4: Construct a super-resolution network model containing a depth-indexed dynamic sparse convolutional layer. In this model, establish a step-matrix mapping relationship between depth intervals and dilation rates. From a preset depth-dilation rate lookup table, use the depth value of the sparse depth projection map as an index to retrieve the corresponding dilation rate. Then, dynamically load convolutional kernel weights that match the dilation rate from the shared weight pool to achieve adaptive feature sampling based on physical depth.

[0043] The specific operating mechanism of the depth-indexed dynamic sparse convolutional layer is as follows: a set of basic convolutional kernel parameters is pre-set; a step-wise mapping relationship between the depth interval and the dilation rate within the high-interest reconstruction region is established: the dilation rate gradually increases as the depth increases; during network inference, the dilation rate at the current position is determined by using the depth value corresponding to each position of the feature map as an index; based on the determined dilation rate, non-continuous spatial sampling is performed on the basic convolutional kernel parameter set to reconstruct an equivalent convolutional kernel that fits the current depth, and the equivalent convolutional kernel is used to perform convolution operations on the feature map.

[0044] In this embodiment, the step mapping relationship between the depth range of the high-concern reconstruction region and the dilation rate is established as follows:

[0045] Medium distance range (e.g., 20m) 50m): mapped to expansion rate ;

[0046] Long distance range (e.g., >50m): mapped to expansion rate (Large-diameter convolution, focusing on the outline).

[0047] Step S5: Input the image to be reconstructed into the super-resolution network model, use computing power to allocate masks to control the inference path, only call the dilation rate matching its depth for feature extraction in the high-interest reconstruction area, and perform downsampling or interpolation processing on the non-high-interest reconstruction area, and then output the final image.

[0048] If computing power allocation mask If the value is marked as "0", the network skips the convolution calculation and directly outputs zero or performs bilinear interpolation, which can save a lot of computing power.

[0049] If computing power allocation mask The mark is "1", in depth value For indexing, retrieve the dilation rate from the depth-dilation rate lookup table. Then, matching convolutional kernel weights are loaded from the shared weight pool, feature aggregation is performed, and the final output image is expressed as:

[0050]

[0051] in, The final image in spatial coordinates Pixel feature response at the location; Represents the computing power allocation mask in spatial coordinates The binary gating coefficient at the location, if the spatial coordinates If the area is a high-priority redevelopment zone, then If spatial coordinates If the area is not a high-priority redevelopment zone, then ; This represents the set of discrete spatial sampling neighborhoods defined by the convolution kernel. This indicates traversing the neighborhood set. The relative offset vector, Indicates and The corresponding convolution kernel weight parameters; The input feature map is the image to be reconstructed if the dynamic sparse convolutional layer is the first layer of the network, and the input feature map is the output of the previous layer if the dynamic sparse convolutional layer is an intermediate layer. Based on spatial coordinates The dilation rate is obtained from the depth value index. This represents the feature values ​​of the sampled points in the input feature map after the dilation step size offset. This is the channel bias term.

[0052] By using the above method, a large number of floating-point operations (FLOPs) can be reduced without reducing the accuracy of sensing small targets at long distances.

[0053] In this embodiment, the super-resolution network model introduces a dynamic dilation rate decoupling training strategy during the forward propagation process: During the training phase, a random perturbation mechanism is introduced to impart scale invariance to the model. In each forward propagation iteration, for the same set of shared convolutional kernel weights, they are no longer fixedly associated with a single dilation rate. Instead, they are randomly sampled from a preset set of dilation rates, and a temporary dynamic mapping is established between the sampled dilation rates and the set of weights. This combination is then used to perform convolution operations on the training images. This strategy aims to break the strong coupling between weight parameters and fixed receptive fields in traditional convolutional neural networks, forcing the convolutional kernel to learn a robust feature representation method, enabling it to extract effective image features during the inference phase regardless of the dilation rate applied (i.e., regardless of the physical distance).

[0054] In this embodiment, the loss function of the super-resolution network model for:

[0055]

[0056] in, For pixel-level reconstruction loss, For balance coefficient, This is a perceptual feature loss based on a pre-trained VGG network.

[0057] Parameter Update: The parameters in the shared weight pool are synchronously updated using the backpropagation algorithm and the aforementioned loss function. The resulting weight pool exhibits scale invariance, meaning that the same set of parameters can extract effective features under different inflation rates, thus supporting dynamic loading during the inference phase.

[0058] In summary, the image super-resolution reconstruction method according to the above embodiments has the following beneficial effects:

[0059] (1) To address the issues of blurred details in existing lightweight models when dealing with small targets at long distances, and excessive inference latency in high-performance models, this invention proposes a dynamic sparse convolutional layer based on depth indexing. This layer establishes a mapping relationship between depth intervals and dilation rates, and can adaptively adjust the receptive field size according to the physical distance of the target: standard convolution can be used to preserve texture at close range, while convolution with a large dilation rate can be used to enhance contours at long range. Combined with a dynamic dilation rate decoupling training strategy, this invention effectively improves the super-resolution reconstruction clarity of key targets such as traffic signs at a distance of 100 meters without significantly increasing the number of model parameters. Therefore, this invention balances inference speed and reconstruction accuracy.

[0060] (2) To address the problem that traditional convolutional networks perform equal calculations on all pixels of the image, resulting in a large amount of ineffective power consumption in background areas such as the sky and nearby road surfaces, this invention utilizes the depth information of LiDAR to generate a computing power allocation mask. By introducing a spatial sparsity mechanism during network inference, downsampling or interpolation is performed on non-high-interest reconstruction areas, while high-resolution reconstruction is performed only on high-interest reconstruction areas containing distant key targets. This physically guided computing strategy significantly reduces the floating-point operation volume and chip power consumption of the model, achieving efficient allocation of onboard computing resources.

[0061] (3) To address the problem that existing single-vision solutions are prone to misjudging "false fog" in scenarios such as white walls and highly reflective surfaces, leading to incorrect activation of defogging processing and the introduction of noise, this invention constructs a cross-modal conflict arbitration mechanism based on physical features. This mechanism utilizes the physical reflection characteristics of point cloud data as an objective physical criterion to logically verify visual environment features. This mechanism effectively distinguishes between real aerosol scattering and high surface reflectivity, activating enhancement processing only when both visual environment features and physical reflection features indicate the presence of environmental interference. This avoids the waste of computational power and image distortion risks caused by algorithmic misjudgment at the source, reducing the environmental perception misjudgment rate.

[0062] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An image super-resolution reconstruction method, characterized in that, include: Step S1: Obtain the original image captured by the vehicle-mounted camera and the point cloud data of the LiDAR, project the point cloud data onto the pixel coordinate system of the original image, and generate a sparse depth projection map containing depth values ​​and reflection intensity. Step S2: Extract the visual environment features of the original image and the physical reflection features of the point cloud data, and construct a cross-modal conflict arbitration mechanism. If the visual environment features indicate the presence of environmental interference, but the physical reflection features indicate the absence of environmental interference, the current scene is determined to be a false fog scene. No dehazing preprocessing is performed, and the original image is directly output as the image to be reconstructed. If both the visual environment features and the physical reflection features indicate the presence of environmental interference, the current scene is determined to be a real fog scene. Dehazing preprocessing is performed on the original image, and the generated dehazing enhanced image is used as the image to be reconstructed. Step S3: Based on the depth value in the sparse depth projection map and the preset safety concern distance threshold, generate a computing power allocation mask, mark the region with a depth value greater than the distance threshold as a high concern reconstruction region, and perform a morphological expansion operation on the high concern reconstruction region to make the high concern reconstruction region form a physically continuous spatial structure. Step S4: Construct a super-resolution network model containing a depth-indexed dynamic sparse convolutional layer. In this model, establish a step mapping relationship between depth range and dilation rate. From the preset depth-dilation rate lookup table, use the depth value of the sparse depth projection map as an index to retrieve the corresponding dilation rate. Dynamically load the convolutional kernel weights that match the dilation rate from the shared weight pool to achieve adaptive feature sampling based on physical depth. Step S5: Input the image to be reconstructed into the super-resolution network model, use computing power to allocate masks to control the inference path, only call the dilation rate matching its depth for feature extraction in the high-interest reconstruction area, and perform downsampling or interpolation processing on the non-high-interest reconstruction area, and then output the final image.

2. The image super-resolution reconstruction method according to claim 1, characterized in that, In step S1, during the process of projecting the point cloud data onto the pixel coordinate system of the original image, the following equation is satisfied: in, , , The coordinates are three-dimensional coordinates in the lidar coordinate system. , These are the pixel coordinates of the point cloud data projected onto the original image. Depth in camera coordinate system This is the camera's intrinsic parameter matrix. Let be the rotation matrix of the lidar coordinate system relative to the camera coordinate system. Let be the translation vector of the lidar relative to the camera.

3. The image super-resolution reconstruction method according to claim 1, characterized in that, Step S2 specifically includes: The normalized dark channel average intensity of the original image is calculated to obtain the visual fog index. Visual fog index As a feature of the visual environment; The reflection intensity and distance information of point cloud data are extracted, and the spatial attenuation slope of the reflection intensity is fitted by linear regression. Based on the spatial decay slope Calculate the radar fog perception index , will radar fog perception indicators As a physical reflection characteristic; Arbitration Difference Values ​​for Constructing Cross-Modal Conflict Arbitration Mechanisms , ; When the visual fog index Greater than or equal to a preset visual threshold, and When the collision threshold is greater than or equal to the collision threshold, the current scene is determined to be a false fog scene. No defogging preprocessing is performed, and the original image is directly output as the image to be reconstructed. When the visual fog index Greater than or equal to a preset visual threshold, and When the value is less than the conflict threshold, the current scene is determined to be a real foggy scene. Dehazing preprocessing is performed on the original image, and the generated dehazing enhanced image is used as the image to be reconstructed.

4. The image super-resolution reconstruction method according to claim 3, characterized in that, Radar fog perception index Calculated using the following formula: in, The preset sensitivity coefficient, This is the preset baseline attenuation threshold.

5. The image super-resolution reconstruction method according to claim 4, characterized in that, Step S2 also includes: When the visual fog index When the image size is less than the preset visual threshold, the current scene is determined to be a scene without visual interference. No dehazing preprocessing is performed, and the original image is directly output as the image to be reconstructed.

6. The image super-resolution reconstruction method according to claim 1, characterized in that, The final image output by the super-resolution network model satisfies the following equation: in, The final image in spatial coordinates Pixel feature response at the location; Represents the computing power allocation mask in spatial coordinates The binary gating coefficient at the location, if the spatial coordinates If the area is a high-priority redevelopment zone, then If spatial coordinates If the area is not a high-priority redevelopment zone, then ; This represents the set of discrete spatial sampling neighborhoods defined by the convolution kernel. This indicates traversing the neighborhood set. The relative offset vector, Indicates and The corresponding convolution kernel weight parameters, This is the input feature map for the dynamic sparse convolutional layer. Based on spatial coordinates The dilation rate is obtained from the depth value index. This is the channel bias term.

7. The image super-resolution reconstruction method according to claim 1, characterized in that, Loss function of super-resolution network model for: in, For pixel-level reconstruction loss, For balance coefficient, This is a perceptual feature loss based on a pre-trained VGG network.

Images