Hierarchical scanning perception fusion super-resolution enhanced target detection method and system

Abstract

The application discloses a layered scanning perception fusion super-resolution enhancement target detection method and system, and belongs to the technical field of image processing. The method comprises the following steps: performing target detection on an input image after down-sampling to generate a global coarse detection result; generating a saliency map based on the input image, dividing the input image into multiple image patches, calculating a normalized priority probability according to the saliency map and the position information of each image patch, and generating a scanning sequence; inputting the image patches into a sequence encoder in the order of the scanning sequence to obtain an enhanced feature map; identifying key patches with dense target distribution based on the enhanced feature map by using an adaptive threshold strategy, performing super-resolution reconstruction on the key patches, performing target detection again to obtain a local fine detection result, and outputting a final detection result after fusing the local fine detection result with the global coarse detection result. The application only performs super-resolution reconstruction on the key patches with local dense distribution, thereby ensuring the detection accuracy and reducing the calculation cost.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a hierarchical scanning perception fusion super-resolution enhanced target detection method and system. Background Technology

[0002] With the rapid advancements in remote sensing technology and the continuous improvement in the spatial resolution of remote sensing images, the amount of available remote sensing image data has grown exponentially. This development has made it possible to automatically and accurately extract regions of interest from massive amounts of remote sensing image data. Among these, aircraft target detection in remote sensing images has become one of the research focuses in the field of remote sensing target detection due to its broad application prospects and significant strategic value. However, despite the unprecedented observational detail provided by high-resolution remote sensing satellites and advanced imaging systems, aircraft target detection in remote sensing images still faces several fundamental challenges in practical implementation. First, unlike natural scene images, aircraft targets in remote sensing images are usually extremely small, oriented arbitrarily, and exhibit significant inter-class similarity with man-made objects in the background. Furthermore, the spatial distribution of aircraft targets within airport areas is extremely uneven; they tend to cluster densely in areas such as aprons and taxiways, while being sparsely distributed or completely absent in other vast areas. This characteristic makes traditional processing strategies based on uniform slicing or sliding windows inefficient, as they consume significant computational resources processing target-free backgrounds and may truncate targets at slice boundaries. While high-resolution images offer rich detail, they also introduce complex background interference and scale variations, further complicating the detection of aircraft targets in remote sensing images.

[0003] Existing technologies mainly focus on two directions: First, detectors based on Convolutional Neural Networks (CNNs) improve the detection capability of multi-scale targets by introducing mechanisms such as feature pyramids. However, the locality of their convolutional operations limits their efficiency in modeling the global context, making it difficult to achieve a balance between accuracy and speed in ultra-large images. Second, detection frameworks based on Visual Transformers (ViTs), while capable of capturing long-range dependencies through self-attention mechanisms, have computational complexity proportional to the square of the image size. Directly applying them to high-resolution images would result in unbearable computational overhead, limiting their practical deployment. To improve the detection capability of small targets, another technical approach attempts to combine super-resolution (SR) with detection. However, existing combinations have significant drawbacks: the "global super-resolution first, then detection" strategy involves unnecessary detail reconstruction in vast background areas, resulting in a huge waste of computational resources; while local enhancement strategies often rely on fixed heuristic rules or manual thresholds to select regions, lacking sufficient intelligence and failing to dynamically and accurately identify the key areas with densely distributed targets that truly need refinement based on image content.

[0004] In summary, existing technologies for processing high-resolution remote sensing imagery generally face a core contradiction: the difficulty in simultaneously achieving global context modeling, computational efficiency, and enhanced intelligent resource allocation. In particular, they are unable to dynamically and adaptively focus computational resources to address the sparse distribution and high concentration of targets in remote sensing imagery. This invention is proposed precisely to address these shortcomings. Summary of the Invention

[0005] The purpose of this invention is to provide a hierarchical scanning perception fusion super-resolution enhanced target detection method and system to overcome the shortcomings of the prior art.

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

[0007] In a first aspect, the present invention provides a hierarchical scanning perception fusion super-resolution enhanced target detection method, comprising the following steps:

[0008] S1, Perform object detection on the downsampled image of the input image to generate a global coarse detection result, which includes preliminary bounding box suggestions and scene-level context information;

[0009] S2, Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the normalized priority probability of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order;

[0010] S3, according to the scanning sequence with the saliency derivation priority order, the image patch is input into the sequence encoder for processing, thereby ensuring that the highly saliency region is processed earlier and more focusedly, and then the enhanced one-dimensional feature sequence is inversely mapped back to the two-dimensional space to obtain the enhanced feature map;

[0011] S4. Based on the enhanced feature map, a multi-scale priority scoring network is used to identify and select key patches with dense local distribution of the target from the input image using an adaptive threshold strategy. Local super-resolution reconstruction is then performed on the key patches with dense local distribution of the target to generate super-resolution patches and obtain a super-resolution region group.

[0012] S5, Perform target detection on the super-resolution region group to obtain local fine detection results, process the local fine detection results to obtain local fine detection results after coordinate system transformation, and fuse the local fine detection results after coordinate system transformation with the global coarse detection results using a non-maximum suppression algorithm to output the final target detection result.

[0013] Furthermore, S2 specifically includes:

[0014] The input image is divided into a grid corresponding to the processing resolution of the sequence encoder in S3, with each grid cell corresponding to an image patch, resulting in a set of image patches;

[0015] The input image is processed by a lightweight convolutional network to obtain a basic feature map, and the basic feature map is processed by an Inception-like (multi-scale parallel convolution) module containing four parallel branches to obtain multi-scale features. Combined with the semantic prior provided by the global coarse detection results, a saliency map is generated.

[0016] The saliency features corresponding to the position of each image patch in the set of image patches are extracted from the saliency map, and a position embedding is generated based on the grid coordinates of each image patch in the input image. The position embedding encodes the spatial layout information of each image patch.

[0017] The saliency features of each image patch are fused with the location embedding to obtain a joint representation of the image patch. The joint representation is then input into a multi-scale priority scoring network to calculate the priority score of each image patch.

[0018] The priority score of each image patch is normalized to obtain the normalized priority probability;

[0019] The image patches are sorted in descending order of their normalized priority probability to generate the scan sequence with the saliency-derived priority order.

[0020] Furthermore, S3 specifically includes:

[0021] For each image patch in the image patch set, initial local features are extracted through a convolutional backbone and projected onto an embedding space of a preset dimension to obtain the embedding representation of each image patch, forming a two-dimensional feature grid;

[0022] The embedding representation of each image patch in the two-dimensional feature grid is reorganized into a one-dimensional feature sequence according to the scan sequence with saliency-derived priority order;

[0023] The one-dimensional feature sequence is input into a sequence encoder constructed from multiple Mamba blocks. The Mamba blocks recursively update the hidden state along the one-dimensional feature sequence to obtain an enhanced one-dimensional feature sequence. The hidden state fuses the salient features and spatial layout information of the processed image patches to form a contextual representation, and the contextual representation is recursively propagated to subsequent image patches.

[0024] Based on the inverse mapping relationship of the scan sequence with saliency-derived priority order, the enhanced one-dimensional feature sequence is restored to a two-dimensional spatial layout to obtain the enhanced feature map.

[0025] Furthermore, S4 specifically includes:

[0026] Global average pooling is performed on the enhanced feature map within the spatial range corresponding to each image patch to obtain the patch-level feature vector of each image patch;

[0027] To incorporate contextual information from different receptive fields, multi-scale contextual features corresponding to the location of each image patch are extracted from multiple different resolution levels of the feature pyramid and aligned using adaptive pooling.

[0028] The multi-scale context features corresponding to the location of each image patch are concatenated with the patch-level feature vector to obtain the fused features of each image patch;

[0029] The fused features of each image patch are processed by a linear classifier, followed by linear projection and a Sigmoid activation function to generate selection scores for each image patch at different scales.

[0030] According to the preset selection ratio, the selection scores of each image patch at different scales are fused to obtain the final score of each patch. The final scores of each image patch are sorted in descending order, and a threshold is dynamically and adaptively determined. Image patches with a final score higher than the dynamically and adaptively determined threshold are marked as key patches with dense local distribution of the target.

[0031] High-resolution regions corresponding to the densely distributed key patches of the target area are cropped from the input image and used as low-resolution regions to be super-resolution reconstructed.

[0032] The low-resolution region is input into a multi-directional 2D Selective Scan (SS2D) super-resolution network for reconstruction, generating super-resolution patches and obtaining a group of key patch regions after super-resolution.

[0033] Furthermore, the multi-directional SS2D super-resolution network includes shallow convolutional layers, multiple cascaded multi-directional SS2D modules, and an upsampling module;

[0034] The shallow convolutional layer is used to extract initial features from the low-resolution input region to obtain a shallow feature map.

[0035] The multiple cascaded multi-directional SS2D modules process the data sequentially: the input of the first multi-directional SS2D module is the shallow feature map, and the input of each subsequent multi-directional SS2D module is the feature map output by the previous multi-directional SS2D module.

[0036] Each multi-directional SS2D module is used to process the feature map input to the multi-directional SS2D module along multiple preset scanning directions to obtain the feature map output by the multi-directional SS2D module;

[0037] The upsampling module is used to upsample and reconstruct the feature map output by the last multi-directional SS2D module to generate super-resolution patches and obtain a super-resolution region group.

[0038] Furthermore, the multi-directional SS2D module processes the feature map input to the multi-directional SS2D module along multiple preset scanning directions, specifically including:

[0039] The feature maps input to the multi-directional SS2D module are flattened into a one-dimensional sequence according to the scanning order along four directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right.

[0040] The one-dimensional sequence in each direction is processed by the Mamba layer and then reconstructed back into two-dimensional space to obtain the processing result for each direction.

[0041] The processing results from the four directions are fused together to form the feature map output by the multi-directional SS2D module.

[0042] Furthermore, S5 specifically includes:

[0043] Target detection is performed on the super-resolution key patch region group to obtain the local fine detection results of each super-resolution key patch region group;

[0044] The local fine detection results are transformed back to the coordinate system of the input image through the inverse transformation of the crop offset to obtain the local fine detection results after coordinate system transformation.

[0045] The nonmaximum suppression algorithm is used to fuse the global coarse detection results and the local fine detection results after coordinate system transformation, and the final target detection result is output.

[0046] Furthermore, the hierarchical scanning perception fusion super-resolution enhanced target detection method is optimized end-to-end through a composite loss function, which includes a weighted sum of detection loss, hierarchical patch selection loss, reconstruction loss, and saliency consistency loss.

[0047] Secondly, the present invention also provides a hierarchical scanning perception fusion super-resolution enhanced target detection system for implementing the aforementioned hierarchical scanning perception fusion super-resolution enhanced target detection method, comprising:

[0048] Global coarse detection module: used to perform object detection on the downsampled image of the input image and generate global coarse detection results, which include preliminary bounding box suggestions and scene-level context information;

[0049] Adaptive scan generation module: Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the priority score of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order;

[0050] Scanning perception coding module: used to input the image patches into the sequence encoder for processing according to the scanning sequence with the saliency derivation priority order, to obtain an enhanced one-dimensional feature sequence, and to inversely map the enhanced one-dimensional feature sequence back to two-dimensional space to obtain an enhanced feature map;

[0051] Key Patch Selection and Super-Resolution Module: Based on the enhanced feature map, a multi-scale priority scoring network is used to identify and select key patches with dense distribution in the target locality from the input image, and super-resolution reconstruction is performed on the key patches with dense distribution in the target locality to generate super-resolution patches and obtain super-resolution region groups.

[0052] Local detection and fusion module: This module performs target detection on the super-resolution region group to obtain local fine detection results. It processes the local fine detection results to obtain local fine detection results after coordinate system transformation. It then fuses the local fine detection results after coordinate system transformation with the global coarse detection results using a non-maximum suppression algorithm to output the final target detection result.

[0053] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the hierarchical scanning perception fusion super-resolution enhanced target detection method described above.

[0054] Compared with the prior art, the present invention has the following beneficial technical effects:

[0055] This invention provides a hierarchical scanning-sensory fusion super-resolution enhanced target detection method, abandoning the inefficient traditional approach of global super-resolution followed by detection or uniform processing. It innovatively adopts a coarse-to-fine hierarchical processing paradigm. Target detection is performed on a downsampled image of the input image, generating a global coarse detection result. This obtains semantic priors about scene layout and target distribution at a lower computational cost. Key patches with locally dense target distributions are identified and selected from the input image, and super-resolution reconstruction is performed on these key patches to generate super-resolution patches, resulting in a super-resolution region group. Target detection is then performed only on the selected super-resolution region group. This dynamic allocation of computational resources avoids unnecessary and expensive computations in broad background areas, thereby significantly reducing overall computational overhead while maintaining or even improving the accuracy of key target detection. This solves the core contradiction of balancing efficiency and accuracy in high-resolution image processing. This invention can generate a scanning sequence with a significant derived priority order based on the content of the input image, enabling subsequent sequence encoders to process image patches in descending order of normalized priority probability. It can be trained end-to-end, adapt to different scenarios, and eliminate the reliance on manual thresholds or complex multi-stage training. It achieves targeted allocation of computing resources with linear complexity, ensuring that attention is preferentially focused on aircraft clusters and difficult targets, optimizing the efficiency of feature aggregation in terms of information processing order, and realizing intelligent focusing on key areas.

[0056] Specifically, this invention introduces a Mamba block based on a state-space model as a sequence encoder. Combined with the scanning sequence, the features of high-priority image patches are encoded early and stored in a hidden state. This state serves as a dynamically updated context memory, continuously influencing the feature extraction of subsequent image patches during the scanning process. This enables the state-space model to utilize the processed salient region information to assist in understanding subsequent regions. In particular, it enhances the discriminative power of feature representation for small targets and occluded targets that require context support, achieving effective long-range context modeling with linear computational complexity.

[0057] Specifically, this invention designs a multi-directional SS2D super-resolution network for key map patches with dense local distribution of the selected target. The multi-directional SS2D module of the multi-directional SS2D super-resolution network adapts the one-dimensional state space model to the two-dimensional image data through a scanning strategy in four directions, which can better model the two-dimensional spatial structure of the image, thereby more accurately recovering the structured high-frequency details of the aircraft target. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of a hierarchical scanning perception fusion super-resolution enhanced target detection method according to an embodiment of the present invention.

[0059] Figure 2 This is a flowchart illustrating the multi-directional SS2D super-resolution network in an embodiment of the present invention.

[0060] Figure 3 This is a schematic diagram of the adaptive scanning strategy generation process in an embodiment of the present invention. Detailed Implementation

[0061] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary in nature and not restrictive. The present invention uses a Mamba encoder as the core to achieve efficient long-range feature modeling. In the prior art, fixed, predefined scanning paths cannot prioritize dense clusters or semantically ambiguous regions that are most critical for aircraft detection, resulting in suboptimal feature aggregation and wasted computation in open backgrounds. To solve this problem, the present invention introduces a learnable scanning policy network for adaptive region priority ranking. This learnable scanning policy network directly addresses the uneven spatial distribution of aircraft in remote sensing images by transforming input image processing from a static, parallel operation to a dynamic, sequential decision-making process that plans the optimal trajectory. Therefore, when the Mamba encoder processes an image along this path, features from these high-priority regions are encoded preferentially, and their contextual information is propagated more effectively through the model's hidden states, thereby guiding and enhancing feature extraction for subsequent regions. To implement this adaptive scanning strategy, this invention designs a specialized mechanism consisting of two cascaded components: (i) a lightweight saliency generator that quickly highlights candidate regions based on global cues; and (ii) a learnable scanning policy network that transforms the saliency map and spatial layout into a scanning sequence with a saliency-derived priority order.

[0062] See Figure 1 This invention provides a hierarchical scanning perception fusion super-resolution enhanced target detection method, comprising the following steps:

[0063] S1, Perform object detection on the downsampled image of the input image to generate a global coarse detection result, which includes preliminary bounding box suggestions and scene-level context information;

[0064] In a more specific embodiment provided by the present invention, the input image downsampled image Object detection is performed by a global coarse detector, which generates a global coarse detection result. This result includes preliminary bounding box suggestions and scene-level context information. A standard detection backbone network is used to generate the preliminary bounding box suggestions. And establish scene-level context information.

[0065] It should be noted that S1 quickly obtains the coarse distribution of targets and scene layout in the input image with low computational cost, providing semantic priors for subsequent processing. By performing detection on the downsampled image, the computational cost of the initial processing is significantly reduced, while preserving the contextual information of the global scene.

[0066] S2, Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the normalized priority probability of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order;

[0067] In a more specific embodiment provided by the present invention, S2 specifically includes:

[0068] The input image The image is divided into grids corresponding to the processing resolution of the S3 sequence encoder. Each grid cell corresponds to an image patch, and the image patches are non-overlapping image patches.

[0069] The input image is processed using a lightweight convolutional network. The data is processed, and a saliency map is generated using the semantic priors provided by the global coarse detection results. saliency plot Used to highlight dense areas where aircraft targets may be present; input image It is divided into a grid corresponding to the processing resolution of the sequence encoder, each grid cell Each image tile corresponds to a grid, where N is the total number of image tiles. The image tiles form an H×W grid layout, where H and W are the number of image tiles in the height and width directions of the grid, respectively. This represents the coordinate position of an image patch within a grid; where the lightweight convolutional network is a lightweight saliency generator.

[0070] In a more specific embodiment of the present invention, a saliency map is generated. The process is as follows:

[0071] The input image is processed by a compact convolutional backbone. The basic feature map is obtained through processing.

[0072] The basic feature map is processed by an Inception-like module containing four parallel branches to capture multi-scale contextual information and obtain multi-scale features.

[0073] The multi-scale features are fused through a stitching operation, and then channel compression and feature integration are performed through a 1×1 convolution to obtain the fused feature map.

[0074] The fused feature map is converted into a dual-channel activation map and normalized using Softmax of the channel dimension to generate a saliency map. saliency plot Each element value in the quantization is a pixel. The confidence score for an aircraft target indicates the likelihood of its presence at that location and in its neighborhood. The core objective of this probability estimation is to highlight densely distributed areas of aircraft targets. When multiple aircraft are densely packed in a local area (such as an apron), their individual probabilities spatially superimpose, forming high-response "hotspot" clusters on the saliency map. This clustering effect allows the algorithm to macroscopically identify areas requiring focused attention, rather than processing each pixel in isolation.

[0075] For each grid cell From the saliency graph The salient features corresponding to the position of each image patch in the image patch set are extracted, and a position embedding is generated based on the grid coordinates of each image patch in the input image. The position embedding encodes the spatial layout information of each image patch.

[0076] The saliency features of each image patch are fused with the location embedding to obtain a joint representation of the image patch. The joint representation The input is fed into a multi-scale priority scoring network to calculate the normalized priority probability of each image patch; the multi-scale priority scoring network is the learnable scanning strategy network in this invention;

[0077] Specifically, a lightweight multilayer perceptron (MLP) is used to assign a single scalar priority score to each image patch:

[0078] ;

[0079] in, It is a three-layer MLP with ReLU activation function. This indicates its learnable parameters.

[0080] The priority scores of all image patches are normalized to obtain normalized priority probabilities. :

[0081]

[0082] in, This indicates the index of the currently calculated image patch within an M×N set of image patches. This represents the index variable used to iterate through all image tiles. The original priority score of the i-th image patch is represented by the scalar score calculated by the multilayer perceptron (MLP), which indicates the importance of the image patch. This represents the original priority score of the j-th image patch.

[0083] The image patches are sorted in descending order of their normalized priority probability to generate the scan sequence with the saliency-derived priority order.

[0084] Specifically, scan sequences with a significant derived priority order. The expression is:

[0085]

[0086] in, This indicates the number of all image tiles in the grid. The grid coordinates of the first image patch being processed. The grid coordinates of the second image patch being processed. Let T be the grid coordinates of the T-th image patch being processed.

[0087] This prioritization strategy inherently encourages spatial coherence, as aircraft typically cluster in specific areas. Therefore, image patches with salient features often form contiguous blocks, generating a scan sequence that creates a scan path that naturally focuses first on areas of high aircraft density before moving to other areas.

[0088] S3, according to the scanning sequence with the significant derivation priority order, the image patch is input into the sequence encoder for processing to obtain the enhanced one-dimensional feature sequence, and the enhanced one-dimensional feature sequence is inversely mapped back to the two-dimensional space to obtain the enhanced feature map;

[0089] It should be noted that, in a more specific embodiment provided by the present invention, for each image patch in the image patch set, initial local features are extracted through a convolutional backbone and projected onto an embedding space of a preset dimension to obtain an embedding representation of each image patch. All embedding representations form a two-dimensional feature grid. In this embodiment, the preset dimension can be D-dimensional, that is, the initial local features are projected onto a D-dimensional embedding space to form a two-dimensional feature grid. , where D is the embedding dimension.

[0090] The embedded representations of each image patch in the two-dimensional feature grid are reorganized into a one-dimensional feature sequence according to the scan sequence with the saliency-derived priority order;

[0091] In a more specific embodiment provided by the present invention, a two-dimensional feature mesh is used. The embedding representation of each image patch in the sequence is according to the scan sequence. Reconverted into one-dimensional feature sequences ,in, For the embedding representation of the Nth processed image patch, the one-dimensional feature sequence is a sequence with saliency derivation priority;

[0092] The one-dimensional feature sequence is input into a sequence encoder constructed from multiple Mamba blocks. The Mamba blocks are based on the State Space Model (SSM) and recursively update the hidden state along the one-dimensional feature sequence to obtain an enhanced one-dimensional feature sequence.

[0093] In a more specific embodiment provided by the present invention, for the t-th element in the one-dimensional feature sequence The Mamba processing procedure is represented as follows:

[0094]

[0095] in, The current hidden state. This is the hidden state from the previous moment. The output features at the current time. and The discretization parameters are calculated using the following formula:

[0096]

[0097]

[0098] in, It is input Relevant learnable parameters, It is the identity matrix;

[0099] After stacking multiple Mamba blocks, an enhanced one-dimensional feature sequence is obtained. ;

[0100] The hidden state at the current moment The feature sequence is recursively updated along the one-dimensional feature sequence, fusing the salient features and spatial layout information of the processed image patches. Features from high-priority image patches are encoded into the hidden state early on. This fused contextual prior propagates through the recursive state, implicitly influencing the feature transformation of subsequent low-priority image patches, so that semantic information flows from the salient region to its surrounding region.

[0101] Based on the inverse mapping relationship of the scan sequence with saliency-derived priority order, the enhanced one-dimensional feature sequence is restored to a two-dimensional spatial layout to obtain the enhanced feature map. In enhanced feature maps In this model, features from high-priority regions gain richer contextual information through early encoding and hidden state propagation, thereby enhancing the feature representation capabilities for small targets and occluded targets.

[0102] S4, based on the enhanced feature map The key patches with dense distribution in the local area of ​​the target are identified and selected from the input image, and super-resolution reconstruction is performed on the key patches with dense distribution in the local area of ​​the target to obtain super-resolution patches.

[0103] In a more specific embodiment provided by the present invention, the enhanced feature map Global average pooling is performed within the spatial range corresponding to each image patch to obtain the patch-level feature vector for each image patch. ;

[0104] To incorporate contextual information from different receptive fields, multi-scale contextual features corresponding to the location of each image patch are extracted from multiple resolution levels of the feature pyramid, denoted as . The multi-scale context features are aligned to the same dimension through adaptive pooling, and the multi-scale context features corresponding to each image patch position are compared with the patch-level feature vector. The images are then stitched together to obtain the fusion features of each image patch.

[0105] The fused features of each image patch are processed by a linear classifier, followed by linear projection and a Sigmoid activation function to generate a selection score for each image patch at scale s. Where i is an image patch and s represents the scale;

[0106] Select the proportion according to the preset target. An adaptive non-fixed threshold strategy is employed to identify key regions requiring high-resolution refinement; the selection scores of each image patch at different scales are fused to obtain the final score for each patch. The final score for each image patch. Sort in descending order, the dynamic adaptive threshold T is set to the [number]th [level]. The value of the bit, where It represents the total number of image tiles in the grid;

[0107] All satisfied Image patches are marked as key patches with densely distributed targets, and the coordinate set of these key patches is obtained. , where K is the number of key patches with dense target distribution; for unselected image patches, bilinear interpolation is applied for upsampling to balance computational efficiency and detection accuracy;

[0108] From input image Extract the set of key tile coordinates that are densely distributed with the target from the middle. The high-resolution region corresponding to each element is used as the low-resolution region to be super-resolution reconstructed. To obtain the set of key regions , where M is the number of key regions;

[0109] Each low-resolution region The super-resolution patches are obtained by reconstructing the data using a multi-directional SS2D super-resolution network. The multi-directional SS2D super-resolution network is a multi-directional two-dimensional selective scanning super-resolution network.

[0110] In a more specific embodiment provided by the present invention, the multi-directional SS2D super-resolution network includes shallow convolutional layers, multiple cascaded multi-directional SS2D modules, and an upsampling module. Figure 2 This is a schematic diagram of the multi-directional SS2D super-resolution network of the present invention.

[0111] Specifically, the shallow convolutional layer is used to extract initial features from the low-resolution region of the input, resulting in a shallow feature map. ,in, It is a shallow convolutional layer;

[0112] Multiple cascaded multi-directional SS2D modules are processed sequentially. In this embodiment, L cascaded multi-directional SS2D modules are used to gradually extract deep layered features:

[0113]

[0114] in, This represents the feature map output by the Lth SS2DBlock, where SS2DBlock represents the SS2D module.

[0115] Figure 2 The text details the internal structure of a multi-directional SS2D module. This module processes feature maps input to it along multiple preset scanning directions, specifically including:

[0116] Along the four directions of top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right, for each direction, the feature map input to the multi-directional SS2D module is flattened into a one-dimensional sequence according to the scanning order. The one-dimensional sequence of each direction is processed by the Mamba layer and then reconstructed back into two-dimensional space to obtain the processing result corresponding to each direction. The processing results of the four directions are fused together to obtain the feature map output by the multi-directional SS2D module.

[0117] Specifically, the feature map input to the multi-directional SS2D module first passes through an embedding layer, mapping each pixel position in the feature map to a continuous high-dimensional vector representation, completing the transformation from the feature space to the sequence representation space. The embedded feature representation is processed along four preset directions: top-left to bottom-right, bottom-right to top-left, top-right to bottom-left, and bottom-left to top-right. For each direction, the embedded feature map is unfolded into a one-dimensional sequence according to the corresponding scanning order. The one-dimensional sequence for each direction is first transformed through a linear layer to obtain the feature dimension, increasing the feature dimension to the scale required for processing within the Mamba block; then, sequence modeling is performed by the specific Mamba block described in S2 to capture long-distance dependencies in the sequence; finally, the processed sequence is reconstructed back into a two-dimensional feature map to obtain the output for that direction. The feature maps obtained from the four directions are merged through a fusion operation to serve as the output feature map of the multi-directional SS2D module.

[0118] The upsampling module, based on pixel rearrangement operations, reconstructs the feature map output by the last multi-directional SS2D module to generate super-resolution patches. After obtaining the super-resolution region group ;

[0119] Depend on Figure 2 As can be seen, this multi-directional SS2D super-resolution network effectively captures long-distance dependencies in images and improves the quality of super-resolution reconstruction through shallow convolutional layers for embedding representation, linear dimensionality transformation in multi-directional scanning, and sequence modeling in Mamba layers. Specifically, the embedding layers complete the mapping from discrete pixels to a continuous feature space, while the linear layers perform dimensionality adaptation in each processing direction.

[0120] S5, for the super-resolution dense patch region group Target detection is performed and processed by a local fine detector to obtain local fine detection results. This local fine detector shares the same architecture design as the global detector in S1, but operates on the enhanced high-resolution input.

[0121] In a more specific embodiment provided by the present invention, the local fine detector outputs a set of local fine detection results for each super-resolution region j. .

[0122] The results of the local fine detection The image is converted back to the input image by inverse cropping offset transformation. The coordinate system was changed to obtain the local fine detection results after the coordinate system transformation. ;

[0123] The nonmaximum suppression algorithm is used to evaluate the global coarse detection results and the local fine detection results after coordinate system transformation. The nonmaximum suppression algorithm is used for fusion, and the final target detection result is output.

[0124] In a more specific embodiment of the present invention, the hierarchical scanning perception fusion super-resolution enhanced target detection method is optimized end-to-end using a composite loss function. The composite loss function comprises a weighted sum of detection loss, hierarchical patch selection loss, reconstruction loss, and saliency consistency loss.

[0125] The overall objective function of the composite loss function is expressed as:

[0126]

[0127] in, These are the hyperparameters for detection loss, hierarchical patch selection loss, reconstruction loss, and saliency consistency loss, respectively.

[0128] The detection loss simultaneously supervises the generation process of the global coarse detection result and the local fine detection result, following the standard Faster R-CNN formula, and is combined with a classification loss for the target category. and regression loss used for bounding box coordinates :

[0129] ;

[0130] in, It is the predicted class probability of the i-th anchor box. It is the true class probability of the i-th anchor box. These are the predicted bounding box parameters of the i-th anchor box. These are the actual bounding box parameters of the i-th anchor box. It is the normalization term for the classification task. It is the normalization term for the regression task.

[0131] To ensure that the multi-scale priority scoring network in S4 can reliably identify densely populated aircraft regions requiring refinement, a hierarchical binary cross-entropy loss is employed; if an image patch is at a certain scale... The intersection-over-union (IoU) ratio with any real bounding box exceeds a threshold. Then its true label Set to 1 otherwise to 0. The hierarchical binary cross-entropy loss is calculated at three pyramid levels:

[0132]

[0133] in, It is a scale The predicted choice probability, where BCE represents the binary cross-entropy. These are scale-specific weights.

[0134] For the super-resolution network in S4, a multi-scale priority scoring network with pixel-level loss and perceptual loss is used:

[0135]

[0136] in, This represents the k-th super-resolution reconstructed image patch. For the k-th real high-resolution image patch, As a feature, It is a weighting factor for perceived loss.

[0137] To guide the lightweight convolutional network in S2 to highlight regions that may contain aircraft, a consistency loss with the approximate true saliency map is introduced; the approximate true saliency map Generated by dilating all real bounding boxes:

[0138]

[0139] in, This is the significance plot predicted in S2. Binary cross-entropy loss combined with the Sigmoid activation function.

[0140] To further verify the technical effectiveness of the proposed Hierarchical Saliency-Guided Super-Resolution Framework for Robust Aircraft Detection (HSRDet), this embodiment conducted quantitative and qualitative experiments on a publicly available large-scale remote sensing image dataset and compared it with existing mainstream target detection methods.

[0141] In this embodiment, target detection experiments were conducted on the DIOR dataset, a large-scale benchmark dataset for general target detection in optical remote sensing images. The DIOR dataset contains 23,463 images, each 800×800 pixels in size, with ground sampling distances ranging from 0.5 to 30 meters. The dataset includes 20 target categories and a total of 192,472 labeled examples, all labeled using horizontal bounding boxes. This invention focuses on aircraft target detection; therefore, the experimental evaluation primarily targets the aircraft category in the DIOR dataset. To establish a comprehensive performance benchmark, the HSRDet of this invention is compared with six state-of-the-art detectors representing different architectural paradigms. In the CNN-based detectors, this embodiment selected the single-stage RetinaNet and the two-stage Faster Region-based Convolutional Neural Network (Faster RCNN); in the attention-based detectors, the Detection Transformer (DETR) and the Region of Interest Transformer (RoI Transformer) were compared; in the super-resolution enhancement detectors, the Global-Local Super-Resolution Attention Network (GLSAN) and Super Resolution Assisted YOLO Object Detection (SuperYOLO) were compared; all comparison models were retrained from scratch using their official implementations under the same training protocol and data partitioning to ensure fair and reproducible evaluation.

[0142] To ensure a fair comparison, following standard protocols, performance is evaluated using the horizontal bounding rectangle of the Common Objects in Context (COCO) dataset, specifically including the following key metrics: and The positioning quality was measured at IOU thresholds of 0.50 and 0.75, respectively; scale-sensitive index and Used to evaluate small targets (area) Pixels), Medium Target ( area ) and large goals (area) These metrics collectively provide a rigorous and comprehensive assessment of the classification and localization performance of aircraft targets at different scales.

[0143] Table 1 shows the detection accuracy of different target detection methods on the DIOR dataset in this embodiment. Experimental results show that the HSRDet model proposed in this invention achieves the best performance in all evaluation metrics.

[0144] Table 1. Comparison of aircraft detection accuracy of different methods on the DIOR dataset.

[0145]

[0146] Specifically:

[0147] Overall detection accuracy: This invention achieves the highest AP (Accuracy Rate). 50 (95.5%) and AP 75 (78.4%), especially in the case of high-precision positioning AP 75 Compared to the closest comparable method, GLSAN (67.7%), the performance improvement is 15.8%, demonstrating a significant advantage. This verifies the effectiveness of the "coarse-to-fine" hierarchical processing paradigm of this invention in accurately locating aircraft targets at high IoU thresholds.

[0148] Small target detection capability: This invention excels particularly in small-scale aircraft detection, AP S The accuracy rate reached 43.3%, significantly outperforming all baseline methods, including GLSAN (38.9%) and RoI Transformer (24.2%). This verifies that the present invention can accurately identify densely populated regions of small targets through the overall S2 process and multi-scale priority scoring network, and effectively recover discriminative detail features of densely distributed target regions through a multi-directional SS2D super-resolution network.

[0149] Medium and large target detection: This invention also achieves optimal results for medium and large aircraft, AP M 65.1%, AP L With a resolution of 77.9%, it demonstrates robustness across scale variations. This indicates that the Mamba encoder can effectively model long-range contextual information, enhancing its ability to represent features of multi-scale targets.

[0150] To evaluate the contributions of each component in the HSRDeet framework of this invention, a series of ablation experiments were also conducted on the DIOR validation set in this embodiment. All experiments used the same Faster R-CNN with Residual Network with 50 layers (ResNet-50) as the global coarse detector to ensure fair comparisons under consistent training data and hyperparameter settings. Table 2 shows the target detection performance of the Mamba encoder under different scanning strategies on the DIOR validation set.

[0151] Table 2. Impact of different scanning strategies on the target detection performance of the Mamba encoder.

[0152]

[0153] Experimental results demonstrate that the learnable scanning strategy of this invention achieves optimal performance across all evaluation metrics. Most notably, it improves APS by 34.9% (0.433 vs. 0.321) compared to the raster scanning baseline. This improvement highlights the core function of the learnable scanning strategy: dynamically identifying and prioritizing densely populated or semantically ambiguous regions at the beginning of a scan sequence with a saliency-derived priority order. By encoding these saliency features early, Mamba blocks establish a more informative contextual prior in their cyclic hidden states. This more informative contextual prior information propagates effectively through sequential scans, guiding and enhancing the feature representations of subsequent, less salient image patches, which is particularly beneficial for distinguishing challenging small-scale targets.

[0154] To visually verify the impact of the learnable scanning strategy of this invention on its key intermediate outputs and the resulting scanning sequence with a significant derived priority order, Figure 3 This demonstrates the complete decomposition of a representative airport scene from the DIOR dataset.

[0155] Figure 3 (a) is the input image, a representative airport scene from the DIOR dataset, containing aircraft targets of different scales, densely distributed in the tarmac area. Figure 3 In (b), the lightweight saliency generator generates the corresponding saliency map. The purple / light areas represent areas where the aircraft is likely to be present, and the blue / dark areas represent background areas. It successfully highlights areas where the aircraft is likely to be present and effectively suppresses broad background areas such as runways and grass. Figure 3 In the middle (c), the saliency map is a grid, which aggregates the pixel-by-pixel saliency maps into a grid form corresponding to the division of image patches. Each cell represents a corresponding non-overlapping image patch, forming the basis of the scan sequence with a saliency-derived priority order. Figure 3 In the diagram (d), the scan probability distribution is calculated directly from the grid saliency map using a multi-scale priority scoring network. The cell with the highest probability corresponds to the most crowded aircraft parking area, indicating that the strategy of this invention successfully transforms semantic saliency into meaningful processing priority. Figure 3 The middle (e) shows the scanning path (first 40 steps), which is a visualization of the first 40 steps of the scanning sequence with a significant derived priority order generated based on the scanning probability distribution. It focuses on high-probability clusters (aircraft-dense areas) and processes image patches in order of priority from high to low. Figure 3 In the middle (f), the complete scan path diagram is generated.

[0156] In contrast, traditional raster scanning follows a fixed, predetermined order (from left to right, from top to bottom), processing large areas of empty runways and grass before handling key aircraft clusters. This rigid sequence fails to utilize the contextual priors established by saliency maps, resulting in computational inefficiency in informationless regions and delaying feature encoding of key targets. More importantly, isolated small aircraft are processed later in the raster scan sequence, hindering the model from using early contextual propagation to resolve such ambiguities.

[0157] This visualization verifies the functionality of the learnable scanning strategy of the present invention as an efficient intelligent scheduler: by planning a processing path and dynamically allocating computational attention, it first processes dense target clusters and semantically challenging small targets, while delaying or skipping vast empty backgrounds, thereby significantly improving computational efficiency while ensuring detection accuracy.

[0158] In a more specific embodiment provided by the present invention, the present invention provides a hierarchical scanning perception fusion super-resolution enhanced target detection system for implementing a hierarchical scanning perception fusion super-resolution enhanced target detection method, comprising:

[0159] Global coarse detection module: used to perform object detection on the downsampled image of the input image and generate global coarse detection results, which include preliminary bounding box suggestions and scene-level context information;

[0160] Adaptive scan generation module: Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the normalized priority probability of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order;

[0161] Scanning perception coding module: used to input the image patches into the sequence encoder for processing according to the scanning sequence with the saliency derivation priority order, to obtain an enhanced one-dimensional feature sequence, and to inversely map the enhanced one-dimensional feature sequence back to two-dimensional space to obtain an enhanced feature map;

[0162] Key Patch Selection and Super-Resolution Module: Based on the enhanced feature map, a multi-scale priority scoring network is used to identify and select key patches with dense distribution in the target locality from the input image, and super-resolution reconstruction is performed on the key patches with dense distribution in the target locality to generate super-resolution patches and obtain super-resolution region groups.

[0163] Local detection and fusion module: This module performs target detection on the super-resolution region group to obtain local fine detection results. It processes the local fine detection results to obtain local fine detection results after coordinate system transformation. It then fuses the local fine detection results after coordinate system transformation with the global coarse detection results using a non-maximum suppression algorithm to output the final target detection result.

[0164] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a hierarchical scanning perception fusion super-resolution enhanced target detection method.

[0165] In another embodiment of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a hierarchical scanning perception fusion super-resolution enhanced target detection method.

[0166] This invention is described based on flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to specific embodiments. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowcharts and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art can modify the technical solutions adopted in the above embodiments, or make equivalent substitutions for some of the technical features; and all such modifications and substitutions should fall within the protection scope of the present invention.

Claims

1. A hierarchical scanning perception fusion super-resolution enhancement target detection method, characterized in that, Includes the following steps: S1, Perform object detection on the downsampled image of the input image to generate a global coarse detection result, which includes preliminary bounding box suggestions and scene-level context information; S2, Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the normalized priority probability of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order; S3, according to the scanning sequence with the significant derivation priority order, the image patch is input into the sequence encoder for processing to obtain the enhanced one-dimensional feature sequence, and the enhanced one-dimensional feature sequence is inversely mapped back to the two-dimensional space to obtain the enhanced feature map; S4. Based on the enhanced feature map, the target locally densely distributed key map patches are identified and selected from the input image through a multi-scale priority scoring network and an adaptive threshold strategy. Super-resolution reconstruction is then performed on the target locally densely distributed key map patches to generate a super-resolution key map patch region group. S4 specifically includes: Global average pooling is performed on the enhanced feature map within the spatial range corresponding to each image patch to obtain the patch-level feature vector of each image patch; Multi-scale contextual features corresponding to the location of each image patch are extracted from multiple different resolution levels of the feature pyramid and aligned using adaptive pooling. The multi-scale context features corresponding to the location of each image patch are concatenated with the patch-level feature vector to obtain the fused features of each image patch; The fused features of each image patch are processed by a linear classifier, followed by linear projection and a Sigmoid activation function to generate selection scores for each image patch at different scales. According to the preset selection ratio, the selection scores of each image patch at different scales are fused to obtain the final score of each patch. The final scores of each image patch are sorted in descending order, and a threshold is dynamically and adaptively determined. Image patches with a final score higher than the dynamically and adaptively determined threshold are marked as key patches with dense local distribution of the target. High-resolution regions corresponding to the densely distributed key patches of the target area are cropped from the input image and used as low-resolution regions to be super-resolution reconstructed. The low-resolution region is input into a multi-directional SS2D super-resolution network for reconstruction. The multi-directional SS2D super-resolution network includes shallow convolutional layers, multiple cascaded multi-directional SS2D modules, and an upsampling module; The shallow convolutional layer is used to extract initial features from the low-resolution input region to obtain a shallow feature map. The multiple cascaded multi-directional SS2D modules process the data sequentially: the input of the first multi-directional SS2D module is the shallow feature map, and the input of each subsequent multi-directional SS2D module is the feature map output by the previous multi-directional SS2D module. Each multi-directional SS2D module is used to process the feature map input to the multi-directional SS2D module along multiple preset scanning directions to obtain the feature map output by the multi-directional SS2D module; The upsampling module is used to upsample and reconstruct the feature map output by the last multi-directional SS2D module to generate super-resolution map patches and obtain a group of super-resolution key map patch regions. S5, Perform target detection on the super-resolution key map region group to obtain local fine detection results, process the local fine detection results to obtain local fine detection results after coordinate system transformation, and fuse the local fine detection results after coordinate system transformation with the global coarse detection results using a non-maximum suppression algorithm to output the final target detection results.

2. The hierarchical scanning sensing fusion super-resolution enhanced target detection method according to claim 1, characterized in that, S2 specifically includes: The input image is divided into a grid corresponding to the processing resolution of the S3 sequence encoder, with each grid cell corresponding to an image patch, resulting in a set of image patches; The input image is processed by a lightweight convolutional network to obtain a basic feature map, and the basic feature map is processed by an Inception-like module containing four parallel branches to obtain multi-scale features. Combined with the semantic prior provided by the global coarse detection results, a saliency map is generated. The saliency features corresponding to the position of each image patch in the set of image patches are extracted from the saliency map, and a position embedding is generated based on the grid coordinates of each image patch in the input image. The position embedding encodes the spatial layout information of each image patch. The saliency features of each image patch are fused with the location embedding to obtain a joint representation of the image patch. The joint representation is then input into a multi-scale priority scoring network to calculate the priority score of each image patch. The priority score of each image patch is normalized to obtain the normalized priority probability; The image patches are sorted in descending order of their normalized priority probability to generate the scan sequence with the saliency-derived priority order.

3. The hierarchical scanning sensing fusion super-resolution enhanced target detection method according to claim 2, characterized in that, S3 specifically includes: For each image patch in the image patch set, initial local features are extracted through a convolutional backbone and projected onto an embedding space of a preset dimension to obtain the embedding representation of each image patch, forming a two-dimensional feature grid; The embedding representation of each image patch in the two-dimensional feature grid is reorganized into a one-dimensional feature sequence according to the scan sequence with saliency-derived priority order; The one-dimensional feature sequence is input into a sequence encoder constructed from multiple Mamba blocks. The Mamba blocks recursively update the hidden state along the one-dimensional feature sequence to obtain an enhanced one-dimensional feature sequence. The hidden state fuses the salient features and spatial layout information of the processed image patches to form a contextual representation, and the contextual representation is recursively propagated to subsequent image patches. Based on the inverse mapping relationship of the scan sequence with saliency-derived priority order, the enhanced one-dimensional feature sequence is restored to a two-dimensional spatial layout to obtain the enhanced feature map.

4. The hierarchical scanning sensing fusion super-resolution enhanced target detection method according to claim 1, characterized in that, The multi-directional SS2D module processes the feature map input to the multi-directional SS2D module along multiple preset scanning directions, specifically including: The feature maps input to the multi-directional SS2D module are flattened into a one-dimensional sequence according to the scanning order along four directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right. The one-dimensional sequence in each direction is processed by the Mamba layer and then reconstructed back into two-dimensional space to obtain the processing result for each direction. The processing results from the four directions are fused together to form the feature map output by the multi-directional SS2D module.

5. The hierarchical scanning perception fusion super-resolution enhanced target detection method according to claim 1, characterized in that, S5 specifically includes: Target detection is performed on the super-resolution key map region group to obtain local fine detection results for each super-resolution region; The local fine detection results are transformed back to the coordinate system of the input image through the inverse transformation of the crop offset to obtain the local fine detection results after coordinate system transformation. The nonmaximum suppression algorithm is used to fuse the global coarse detection results and the local fine detection results after coordinate system transformation, and the final target detection result is output.

6. The hierarchical scanning sensing fusion super-resolution enhanced target detection method according to claim 1, characterized in that, The hierarchical scanning perception fusion super-resolution enhanced target detection method described above is optimized end-to-end using a composite loss function, which includes a weighted sum of detection loss, hierarchical patch selection loss, reconstruction loss, and saliency consistency loss.

7. A hierarchical scanning perception fusion super-resolution enhanced target detection system, characterized in that, A hierarchical scanning perception fusion super-resolution enhanced target detection method according to any one of claims 1-6 includes: Global coarse detection module: used to perform object detection on the downsampled image of the input image and generate global coarse detection results, which include preliminary bounding box suggestions and scene-level context information; Adaptive scan generation module: Based on the input image and combined with the scene-level context information, generate a saliency map; divide the input image into multiple image patches, calculate the normalized priority probability of each image patch according to the saliency map and the position information of each image patch in the input image, and generate a scan sequence with saliency-derived priority order; Scanning perception coding module: used to input the image patches into the sequence encoder for processing according to the scanning sequence with the saliency derivation priority order, to obtain an enhanced one-dimensional feature sequence, and to inversely map the enhanced one-dimensional feature sequence back to two-dimensional space to obtain an enhanced feature map; Key Patch Selection and Super-Resolution Module: Based on the enhanced feature map, a multi-scale priority scoring network is used to identify and select key patches with dense distribution in the target locality from the input image, and super-resolution reconstruction is performed on the key patches with dense distribution in the target locality to generate super-resolution patches and obtain super-resolution region groups. The key tile selection and super-resolution module specifically includes: Global average pooling is performed on the enhanced feature map within the spatial range corresponding to each image patch to obtain the patch-level feature vector of each image patch; Multi-scale contextual features corresponding to the location of each image patch are extracted from multiple different resolution levels of the feature pyramid and aligned using adaptive pooling. The multi-scale context features corresponding to the location of each image patch are concatenated with the patch-level feature vector to obtain the fused features of each image patch; The fused features of each image patch are processed by a linear classifier, followed by linear projection and a Sigmoid activation function to generate selection scores for each image patch at different scales. According to the preset selection ratio, the selection scores of each image patch at different scales are fused to obtain the final score of each patch. The final scores of each image patch are sorted in descending order, and a threshold is dynamically and adaptively determined. Image patches with a final score higher than the dynamically and adaptively determined threshold are marked as key patches with dense local distribution of the target. High-resolution regions corresponding to the densely distributed key patches of the target area are cropped from the input image and used as low-resolution regions to be super-resolution reconstructed. The low-resolution region is input into a multi-directional SS2D super-resolution network for reconstruction. The multi-directional SS2D super-resolution network includes shallow convolutional layers, multiple cascaded multi-directional SS2D modules, and an upsampling module; The shallow convolutional layer is used to extract initial features from the low-resolution input region to obtain a shallow feature map. The multiple cascaded multi-directional SS2D modules process the data sequentially: the input of the first multi-directional SS2D module is the shallow feature map, and the input of each subsequent multi-directional SS2D module is the feature map output by the previous multi-directional SS2D module. Each multi-directional SS2D module is used to process the feature map input to the multi-directional SS2D module along multiple preset scanning directions to obtain the feature map output by the multi-directional SS2D module; The upsampling module is used to upsample and reconstruct the feature map output by the last multi-directional SS2D module to generate super-resolution map patches and obtain a group of super-resolution key map patch regions. Local detection and fusion module: This module performs target detection on the super-resolution region group to obtain local fine detection results. It processes the local fine detection results to obtain local fine detection results after coordinate system transformation. It then fuses the local fine detection results after coordinate system transformation with the global coarse detection results using a non-maximum suppression algorithm to output the final target detection result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a hierarchical scanning perception fusion super-resolution enhanced target detection method as described in any one of claims 1 to 6.

Images