An end-to-end infrared image super-resolution target detection method and system

Abstract

The application discloses an end-to-end infrared image super-resolution target detection method and system, which comprises the following steps: A1, collecting original infrared images and performing pretreatment; A2, extracting an initial high-resolution depth feature map and a thermal radiation attention weighting map, and then obtaining an artifact-suppressed high-resolution feature map through a cross-scale thermal radiation consistency gating fusion mechanism; A3, performing double-scale feature differentiation refining, spatial information injection, spatial-channel double-dimension feature activation value competition and cross-level gating adaptive fusion to obtain a fused multi-scale detection feature map; A4, performing boundary box regression to obtain a multi-class detection box parameter set, calculating classification confidence and integrating into an original detection result set; and A5, removing low-confidence detection boxes and then removing overlapping detection boxes through non-maximum suppression to obtain final detection results. The application can solve the problems of artifact interference and small target false alarm and missed detection caused by the separation of image super-resolution reconstruction and target detection in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing and target detection technology, and in particular to an end-to-end infrared image super-resolution target detection method and system. Background Technology

[0002] In applications such as vehicle-mounted environmental perception, infrared thermal imaging has gained widespread attention due to its all-weather operation and resistance to adverse weather conditions such as nighttime, heavy fog, rain, and snow. However, limited by its imaging mechanism, infrared images generally suffer from low resolution, weak contrast, blurred edges, and a lack of texture detail. For example, in nighttime or foggy scenes, distant targets often appear as faint bright spots of only a few pixels in infrared images. The shape and contour information of small-scale targets are missing, increasing the difficulty for detection networks to locate targets and making it difficult to distinguish targets from the background. To alleviate these problems, in engineering practice, super-resolution reconstruction is usually performed before target detection in infrared images. This converts low-resolution infrared images into high-resolution images with clearer visual effects, thereby restoring the structural and edge information of the targets.

[0003] Currently, the mainstream approach in automotive systems is to process super-resolution reconstruction and object detection as two independent steps in a separate pipeline. The front-end super-resolution reconstruction network, such as SRCNN and EDSR models based on convolutional neural networks, typically optimizes pixel-level fidelity metrics like mean squared error or peak signal-to-noise ratio to reconstruct a result that approximates the ideal high-resolution image pixel by pixel. However, due to the significant loss of high-frequency components in infrared low-resolution images, the reconstruction network, when compensating for details, is prone to generating false textures and artifacts at target edges or textured regions that do not match the actual scene's thermal radiation signals. These artifacts do not originate from real targets, but visually resemble the grayscale changes at target edges. The back-end object detection network, such as the YOLO series and Faster R-CNN, directly receives the reconstructed high-resolution image and is designed to identify real targets rather than artificial structures introduced during reconstruction. During perception, the detection network struggles to distinguish whether a local response originates from a real target or a reconstruction artifact. On the other hand, when the super-resolution network suppresses too much noise during denoising and smoothing, it may also erase the already weak edge signals of small targets, causing the detection network to fail to extract target features correctly and resulting in missed detections. Since super-resolution reconstruction and downstream detection are designed and trained independently, the super-resolution module is not aware of the specific requirements of the detection task for feature representation, and the detection module can only passively accept the results generated in the preceding process, and cannot impose detection-oriented feedback constraints on the reconstruction process.

[0004] In recent years, deep learning methods based on convolutional neural networks and Transformers have been further introduced into infrared image restoration and detection. Some studies have attempted to loosely couple reconstruction and detection at the feature level, for example, by using a shared shallow feature encoder to serve both super-resolution reconstruction and detection branches simultaneously, or by adding an auxiliary reconstruction loss to the detection network for fine-tuning. Compared to completely separate pipelines, these approaches have improved feature utilization efficiency to some extent, but the optimization objectives of the reconstruction and detection branches still differ. Artificial textures introduced by reconstruction may still be propagated downstream and interfere with detection decisions, especially when facing distant small targets and low-contrast scenes, where false alarms and missed detections caused by artifacts have not been effectively curbed. Summary of the Invention

[0005] In view of this, the present invention aims to provide an end-to-end infrared image super-resolution target detection method and system to solve the problems of artifact interference and false alarms and missed detections of small targets caused by the separation of image super-resolution reconstruction and target detection in traditional methods.

[0006] An end-to-end infrared image super-resolution target detection method, comprising: A1: Acquire the original infrared image and perform preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain the preprocessed high-resolution image and the preprocessed original resolution image; A2: Based on the preprocessed high-resolution image and the preprocessed original resolution image, the initial high-resolution depth feature map and the thermal radiation attention weighted map are extracted sequentially, and then the artifact-suppressed high-resolution feature map is obtained through the cross-scale thermal radiation consistency gating fusion mechanism. A3: Based on the artifact suppression high-resolution feature map and the preprocessed original resolution image, dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition and cross-level gating adaptive fusion are performed sequentially to obtain a fused multi-scale detection feature map. A4: Based on the fused multi-scale detection feature map, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set; A5: Based on the original detection result set, low-confidence detection boxes are removed sequentially, and overlapping detection boxes are removed by non-maximum suppression to obtain the final detection result.

[0007] Furthermore, step A1 also includes: A11: The original infrared image is acquired by the vehicle-mounted infrared detector, and then the noise is suppressed by the wavelet threshold denoising method. The thermal radiation intensity value is mapped to the preset value range by the linear normalization method to obtain the cleaned original resolution infrared image. A12: The original resolution infrared image after cleaning is spatially expanded using bicubic interpolation to generate an initial high-resolution image; then, the original resolution infrared image after cleaning is enhanced with histogram equalization to obtain an original resolution image with enhanced contrast. A13: By establishing a correspondence between the initial high-resolution image and the original resolution image with enhanced contrast through pixel-level spatial coordinate mapping, the alignment and association of multi-scale images are achieved, resulting in a preprocessed high-resolution image and a preprocessed original resolution image.

[0008] Furthermore, step A2 also includes: A21: Deep learning feature extraction is performed on the preprocessed high-resolution image to obtain an initial high-resolution depth feature map. The calculation method is as follows: ; ; ; ; in, This is a shallow convolutional feature map. This is a convolutional layer with a kernel size of 3×3. For the preprocessed high-resolution image, This is a deformation feature diagram. This is a deformable convolution operation. For GELU functions, For layer normalization, For channel attention gating weights, It is the Sigmoid activation function. This is a convolutional layer with a kernel size of 1×1. For global average pooling, Here is the initial high-resolution depth feature map, ⊕ represents element-wise addition, and ⊙ represents the Hadamard product; A22: Extract thermal physical information from the preprocessed original resolution image to obtain a thermal radiation attention-weighted map, calculated as follows: ; ; ; ; in, This is a local thermal radiation contrast map. To take the absolute value, This is the original resolution image after preprocessing. For a 3×3 window average pooling operation, For local statistical enhancement feature maps, This is a channel-level splicing operation. For a 3x3 window, max pooling operation. This is a feature map encoded by thermal radiation. Attention-weighted plot of thermal radiation; A23: Based on the initial high-resolution depth feature map and the thermal radiation attention-weighted map, feature enhancement and super-resolution reconstruction are performed through a cross-scale thermal radiation consistency gated fusion mechanism to obtain an artifact-suppressed high-resolution feature map. The calculation method is as follows: ; ; ; ; in, This is the upsampled thermal radiation attention map. This is a bilinear upsampling operation. For cross-scale fusion of gated graphs To fuse and enhance feature maps, This is a high-resolution feature map for artifact suppression.

[0009] It should be further explained that in actual working conditions such as vehicle-mounted environmental perception, although infrared thermal imaging can work in all weather conditions and is not easily affected by nighttime, heavy fog, rain, snow and other severe weather, it is limited by the imaging mechanism. Infrared images generally have problems such as low resolution, weak contrast, blurred edges and lack of texture details. The target in front often appears as a faint bright spot of a few pixels in the infrared image. The shape and outline information of small-scale targets are missing. The difference in thermal radiation signals between the target and the background is easily drowned out by environmental noise and scattering, making it difficult for the detection network to establish a stable target localization basis, and making it more difficult to distinguish between the target and the background. In step A2 of this invention, super-resolution reconstruction and target detection are no longer treated as two separate processes. Instead, the super-resolution reconstruction network is designed to directly output a detection-oriented high-resolution feature map. Specifically, an initial high-resolution depth feature map is first extracted from the preprocessed high-resolution image. During this process, deformable convolution operations are used to extract deformable feature maps, enabling feature sampling to adaptively fit the infrared target contours that present different shapes due to changes in distance and viewing angle. This avoids the truncation or blurring of small target edges caused by traditional fixed-grid sampling. At the same time, channel attention gating weights are used to filter feature channels, ensuring that feature responses sensitive to thermal radiation targets are preserved while channel responses with greater detection interference are suppressed. Simultaneously, a thermal radiation attention-weighted map is extracted from the preprocessed original resolution image. This process is based on a local thermal radiation contrast map and combines local statistical enhancement feature maps obtained by 3×3 window average pooling and 3×3 window max pooling operations. The thermal radiation encoded feature map encodes the real thermophysical information. Since the real target is in the infrared image... The corresponding thermal radiation distribution has a certain statistical consistency in local areas, while reconstruction artifacts often exhibit isolated anomalous jumps. Therefore, this thermal radiation attention-weighted map can accurately distinguish between real thermal target areas and false structures introduced by reconstruction. Furthermore, by using a cross-scale thermal radiation consistency gating fusion mechanism, the thermal radiation attention-weighted map is upsampled and fused with the initial high-resolution depth feature map. This makes the high-resolution feature enhancement process constrained by the real thermal radiation physical information from the original resolution image. When a region in the high-resolution feature map shows suspected edge enhancement, the cross-scale fusion gating map will refer to the thermal radiation attention-weighted map at the corresponding location for judgment. If the location does not show consistent thermal radiation characteristics in the original resolution image, the fused enhancement feature map will suppress the response intensity of the region, thereby effectively suppressing false textures and artifacts that do not match the real thermal radiation signal. Finally, an artifact suppression high-resolution feature map is generated, so that the super-resolution process is no longer blindly pursuing pixel-level fidelity, but rather generating feature expressions that are more favorable for detection under the guidance of thermal physics laws. Existing technologies typically treat super-resolution reconstruction and object detection as separate pipelines processed in series. The front-end super-resolution network optimizes pixel-level fidelity metrics such as mean squared error or peak signal-to-noise ratio, tending to generate visually sharper images during training. However, in vehicular infrared scenarios, this sharpness often comes at the cost of introducing false textures. These false textures may correspond to the heat reflection edges of road markings, false hotspot outlines of water stains on the road surface, or false structures generated by the reconstruction network overcompensating for weak target edges. The back-end detection network can only passively receive the preceding reconstruction results. Since the super-resolution module is not aware of the specific feature representation requirements of the detection task, and the detection module cannot impose detection-guided feedback constraints on the reconstruction process, the false structures introduced by the reconstruction are passed downstream and interfere with detection decisions, especially when facing distant small targets and low-contrast scenes. Detection networks are prone to misidentifying reconstructed artifacts as pedestrian or vehicle outlines, triggering false alarms. Furthermore, when super-resolution networks perform strong smoothing to suppress noise, they can also erase the already very weak edge bright spots of small targets, leading to missed detections. Compared to existing technologies, this invention deeply couples super-resolution reconstruction and detection requirements at the feature level, making the super-resolution process directly driven by the detected target. Simultaneously, it utilizes a thermal radiation attention mechanism to extract real thermal physical information from the original resolution image to constrain high-resolution feature generation. Since the thermal radiation information in the original resolution image is not upsampled and distorted, the contrast between the thermal target and the background is more reliable. Introducing this as a constraint into the high-resolution feature enhancement process can improve the ability to represent the edge signals of small targets while effectively suppressing the probability of reconstruction artifacts, thereby improving the false alarm and missed detection problems in vehicle-mounted infrared scenarios.

[0010] Furthermore, step A3 also includes: A31: Based on the artifact-suppressed high-resolution feature map and the preprocessed original resolution image, a dual-stream initial feature extraction network is used. Then, directional gradient response gating and local contrast normalization gating are employed to differentially refine the dual-scale features, resulting in a high-resolution refined feature map and an original resolution refined feature map. The calculation method is as follows: ; ; ; ; in, For high-resolution basic feature maps, For convolutional downsampling networks, This is the original resolution base feature map. For high-resolution refined feature maps, It is a two-dimensional convolutional network. To calculate the gradient magnitude using the Sobel operator, LCN refines feature maps to the original resolution. Local contrast normalization; A32: Spatial information is injected based on the high-resolution refined feature map and the original resolution refined feature map to obtain the original resolution compensated feature map. The calculation method is as follows: ; ; ; in, For high-resolution spatial attention maps, This is the original resolution feature map after upsampling. This is a transposed convolution upsampling operation. The feature map is compensated for at the original resolution. A33: Spatial-channel dual-dimensional feature activation value competition is performed on the high-resolution refined feature map to obtain a high-resolution salient response feature map. The calculation method is as follows: ; ; in, For high-resolution spatially significant response distribution, For group normalization operation, This is the temperature scaling factor. High-resolution salient response feature map; A34: The high-resolution salient response feature map and the original resolution-compensated feature map are fused through cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. The calculation method is as follows: ; ; in, For cross-level gated adaptive fusion vectors, To fuse multi-scale detection feature maps.

[0011] It should be further explained that in vehicle-mounted infrared scenes, although the high-resolution features enhanced by the preceding steps are expanded in spatial size, uncertainties from the reconstruction process may still remain in the region corresponding to small targets at a distance, with false textures mixed with real edges. While the original resolution image maintains a high fidelity of thermal radiation information and preserves the real thermal radiation difference between the target and the background, the target itself occupies only a few pixels, and its shape, outline, and spatial location information are insufficient, making it difficult to support the detection network in establishing an effective basis for target localization. If the detection network relies solely on feature input at a single scale, it is difficult to strike a balance between the risk of artifacts and insufficient resolution: using high-resolution features alone can easily misjudge reconstructed artifacts as real target structures and trigger false alarms, while using original resolution features alone may result in missed detections due to the target being too small and lacking expressive power. In step A3 of this invention, to address the shortcomings of the aforementioned single-scale features, a dual-stream complementary fusion architecture for high-resolution and original-resolution features is designed. Specifically, the high-resolution feature map for artifact suppression and the preprocessed original-resolution image are first subjected to basic feature extraction via a convolutional downsampling network to obtain a high-resolution basic feature map and an original-resolution basic feature map, respectively. For the high-resolution basic feature map, differential refinement is performed using directional gradient response gating. This process uses the Sobel operator to calculate the gradient magnitude to perceive edge structure. Since the real target edge corresponds to the actual thermal radiation signal variation area in the infrared image, its gradient direction and magnitude have a certain continuity, while the false edges generated by reconstructed artifacts often have chaotic gradient directions or magnitudes. Anomalies are observed, therefore, directional gradient response gating can accurately enhance the response of real structural edges while suppressing false edges introduced by reconstruction, resulting in a high-resolution refined feature map. For the original resolution base feature map, local contrast normalization gating is used for differential refinement. This process amplifies the difference between weak thermal signals and the surrounding background through local contrast normalization, enabling real targets occupying only a few pixels to obtain significant feature representation even under the original resolution conditions. This is then processed by a 3×3 convolutional layer and added element-wise to the original resolution base feature map to obtain the original resolution refined feature map. Based on this, a high-resolution spatial attention map guides the original resolution hierarchical compensation generation network to process spatial information. Injection: A high-resolution spatial attention map is generated after processing the high-resolution refined feature map through a 2D convolutional network and layer normalization. This attention map carries the judgment information of the high-resolution branch on the target edge position. The original resolution refined feature map is expanded in spatial size by transposed convolution upsampling to obtain the upsampled original resolution feature map. Then, the high-resolution spatial attention map and the upsampled original resolution feature map are added element-wise after performing a Hadamard product. This allows the target position layout of the original resolution feature map to be guided and constrained by the high-resolution edge information during the expansion of the spatial size, thus obtaining the original resolution compensated feature map. At the same time, spatial-channel dual-dimensional feature activation value competition is performed on the high-resolution refined feature map: after group normalization. The system operates and processes the data using a 2D convolutional network, then divides it by a temperature scaling factor before applying it to a Sigmoid activation function to obtain a high-resolution spatially salient response distribution. This response distribution is then multiplied by a Hadamard product with the high-resolution refined feature map and added element-wise, creating a competitive relationship between activation values ​​in the spatial dimension. This enhances the response at the location of the real target while relatively suppressing the response in the background region, ultimately yielding a high-resolution salient response feature map. Finally, the high-resolution salient response feature map and the original resolution-compensated feature map are fused using cross-level gated adaptive fusion: the two are first concatenated along the channel dimension and then processed by a 2D convolutional network and a Sigmoid activation function to obtain a cross-level gated adaptive fusion vector.The fusion vector adaptively adjusts the contribution weights of the two scale features based on the feature content, enabling the network to automatically select more reliable information sources for different regions. Ultimately, the advantages of the two scales complement each other and the disadvantages suppress each other, generating a fused multi-scale detection feature map. Existing technologies for multi-scale feature fusion typically employ simple upsampling followed by concatenation or direct addition to pass deep semantic information layer by layer to shallower layers. However, in vehicle-mounted infrared small target detection scenarios, these fixed-pattern fusion methods struggle to address the dual challenges of high-resolution artifacts and the scarcity of original resolution information. Existing technologies lack targeted differentiated refinement for the high-resolution and original resolution branches, resulting in unsuppressed artifact noise from the high-resolution branch being directly passed downstream, while the weak target signal from the original resolution branch is diluted by simple upsampling without enhancement. Furthermore, the spatial information injection process in existing technologies lacks cross-scale attention guidance, and the original resolution features easily lose target spatial layout information after upsampling, leading to a decrease in the localization accuracy of small targets. Compared to existing technologies, this invention refines dual-scale features through directional gradient response gating and local contrast normalization gating, enabling the high-resolution branch to suppress artifacts while preserving the true edge structure, and the original resolution branch to maintain thermal radiation fidelity while amplifying weak thermal signals. Furthermore, it guides the original resolution hierarchical compensation through a high-resolution spatial attention map, constraining the spatial expansion process of the original resolution features by high-resolution edge information. Finally, it dynamically adjusts the weights of the two scale features through cross-level gating adaptive fusion vectors, allowing the network to adaptively select more reliable feature sources based on scene content. This improves the stability and accuracy of small target detection in vehicle-mounted infrared scenes and reduces the risk of false alarms and missed detections caused by insufficient single-scale features.

[0012] Furthermore, step A4 also includes: A41: Based on the fusion of multi-scale detection feature maps, target space priors are injected through thermal radiation differential position encoding, and then a detection-specific tensor is constructed by combining dimension alignment mapping. A42: Based on the detection-specific tensor, the thermal radiation-guided YOLO bounding box regression head is used to perform bounding box regression, resulting in a multi-class detection box parameter set; A43: Based on the multi-class detection box parameter set, the thermal deformation corrected detection features are fused with the detection-specific tensor through multi-level thermal contrast-guided classification confidence mapping, the classification confidence is calculated and integrated into the original detection result set.

[0013] Furthermore, step A4 also includes: The height and width of the fused multi-scale detection feature map are concatenated, processed by a 2D grid coordinate generation function, then processed by a 1×1 convolutional layer, divided by a thermal radiation temperature scaling factor, and then processed by a sine function to obtain the thermal radiation differential position code. The fused multi-scale detection feature map is then processed by a 1×1 convolutional layer, and added element-wise with the Hadamard product of the reduced confidence coefficient and the thermal radiation differential position code to obtain the coordinate-enhanced intermediate feature. The coordinate-enhanced intermediate feature is then added element-wise with the fused multi-scale detection feature map, and then processed by layer normalization to obtain the detection-specific tensor. The detection-specific tensor is processed by a 1×1 convolutional layer and then by a Sigmoid activation function to obtain a thermally adaptive modulation coefficient map. The detection-specific tensor and the thermally adaptive modulation coefficient map are then multiplied by a Hadamard layer and processed by a 3×3 convolutional layer to obtain thermally weighted YOLO regression branch intermediate features. The thermally weighted YOLO regression branch intermediate features are then multiplied by a 1×1 convolutional layer and then multiplied by a bounding box parameter scaling factor to obtain a multi-class detection box parameter set. The detection-specific tensor is element-wise summed with the intermediate features of the thermally weighted YOLO regression branch, then processed by a 1×1 convolutional layer, followed by a Sigmoid activation function to obtain the thermal contrast guiding coefficient map. The intermediate features of the thermally weighted YOLO regression branch are processed by a 1×1 convolutional layer, then combined with the thermal contrast guiding coefficient map using a Hadamard product, followed by a Softmax function to obtain the classification confidence matrix. The multi-class detection box parameter set, the classification confidence matrix, and the result of the maximum indexing operation on the classification confidence matrix are integrated to obtain the original detection result set.

[0014] It should be further explained that in dynamic infrared scenes in vehicles, the thermal radiation distribution of the same type of target is uneven due to differences in distance, posture, occlusion, and environmental heat exchange. For example, when driving at night, the thermal radiation intensity of the same pedestrian in the infrared image may differ by several times due to differences in clothing material and distance from the camera. The thermal radiation signal of a distant vehicle's hood, after attenuation, may be similar in intensity to the thermal reflection signal of water stains on the nearby road surface. Traditional detection heads treat infrared targets as visible light targets, ignoring the fact that the thermal radiation intensity of infrared images directly reflects the physical existence of the target. This makes it difficult for the detection head to establish a stable spatial position and classification confidence. Especially when facing low-contrast scenes or partially occluded targets, the detection head lacks effective physical priors to distinguish real targets from similar thermal radiation backgrounds, which can easily lead to positioning errors and classification confusion. In step A4 of this invention, addressing the issue of the detection head's incompatibility with infrared physical characteristics, thermal radiation priors are introduced at three levels: position encoding, bounding box regression, and classification confidence. First, based on the fused multi-scale detection feature map, a thermal radiation differential position encoding is injected into the target space prior: the height and width of the fused multi-scale detection feature map are concatenated, processed by a two-dimensional grid coordinate generation function, then processed by a 1×1 convolutional layer, divided by a thermal radiation temperature scaling factor, and finally processed by a sine function to obtain the thermal radiation differential position encoding. This position encoding combines spatial coordinates with the thermal radiation temperature scaling factor, enabling the detection head to perceive the thermal radiation intensity variation patterns corresponding to different spatial locations, such as the thermal radiation intensity of distant targets. The physical phenomena of weak attenuation of radiation after transmission and strong thermal radiation of nearby targets provide a spatial prior consistent with thermal radiation physics for target localization. Then, the fused multi-scale detection feature map is processed through a 1×1 convolutional layer for dimension alignment mapping, and then element-wise added to the Hadamard product of the reduced confidence coefficient and thermal radiation differential position encoding to obtain coordinate-enhanced intermediate features. These intermediate features are then element-wise added to the fused multi-scale detection feature map and normalized to obtain a detection-specific tensor. Next, based on the detection-specific tensor, a thermal radiation-guided YOLO bounding box regression head is used for bounding box regression: the detection-specific tensor is processed through a 1×1 convolutional layer and then activated by a Sigmoid function. The data processing yields a thermal radiation adaptive modulation coefficient map. This modulation coefficient map generates spatially varying weights based on the thermal radiation characteristics of each location in the detection-specific tensor, enhancing feature expression in regions with strong thermal radiation response and suppressing feature expression in regions with weak thermal radiation response. Then, the detection-specific tensor and the thermal radiation adaptive modulation coefficient map are multiplied by a Hadamard product, and then processed through a 3×3 convolutional layer to obtain thermal radiation-weighted intermediate features of the YOLO regression branch. These intermediate features are then processed through a 1×1 convolutional layer and multiplied by a Hadamard product with the bounding box parameter scaling factor to obtain a multi-class detection box parameter set. Since there is a significant difference in thermal radiation intensity between the real target and the background in the infrared image, this thermal radiation... Adaptive modulation enables bounding box regression to focus more on target regions with realistic thermal radiation responses, reducing localization bias caused by background thermal noise or reconstruction artifacts. Finally, based on the multi-class detection box parameter set, classification confidence is calculated through a multi-level thermal contrast-guided classification confidence mapping: the detection-specific tensor is element-wise summed with the intermediate features of the thermal radiation-weighted YOLO regression branch, then processed by a 1×1 convolutional layer, and finally by a sigmoid activation function to obtain a thermal contrast-guided coefficient map. This coefficient map reflects the degree of thermal radiation difference between the target region and the surrounding background. Real targets usually exhibit relatively consistent thermal contrast characteristics, while artifacts or background noise lack this consistency.The intermediate features of the YOLO regression branch weighted by thermal radiation are then processed by a convolutional layer with a kernel size of 1×1, and then Hadamard product is performed with the thermal contrast guided coefficient map. After processing by the Softmax function, the classification confidence matrix is ​​obtained. Finally, the multi-class detection box parameter set, the classification confidence matrix, and the result of the maximum index operation on the classification confidence matrix are integrated to obtain the original detection result set. Thus, the classification decision is not only based on feature pattern matching, but also constrained by the physical information of thermal contrast, enabling the detection head to output reliable classification results relatively stably when facing infrared scenes with uneven thermal radiation distribution. Existing detection heads typically employ generic position encoding and regression classification methods, treating infrared images as input data homogeneous with visible light images, without specifically designing for the physical characteristics of infrared thermal radiation. In actual vehicle-mounted infrared applications, this generic design makes it difficult for the detection head to handle dynamic changes in thermal radiation distribution. When the target distance changes or the ambient temperature fluctuates, the positioning stability of the detection head decreases. When background interference with similar thermal radiation intensity to the target appears, the classification confidence level tends to fluctuate significantly. Furthermore, existing technologies typically optimize the bounding box regression and classification branches independently. The regression branch does not utilize thermal radiation information to constrain the position and size of the bounding box, and the classification branch does not introduce thermal contrast priors to correct the confidence level, resulting in poor adaptability of the detection head in complex infrared scenes. Limited by limitations in response capabilities; compared to existing technologies, this invention injects thermal radiation spatial priors into the location encoding, enabling the detection head to understand the physical laws governing the variation of thermal radiation with spatial location; by introducing a thermal radiation adaptive modulation coefficient map into the regression branch, the bounding box localization can prioritize responding to regions with real thermal radiation characteristics; by introducing a thermal contrast guiding coefficient map into the classification branch, the classification confidence can reflect the real thermal radiation difference between the target and the background; this end-to-end thermal physical perception design, from location encoding to regression to classification, ensures that the detection decision is consistent with the essential laws of infrared imaging, thereby improving the accuracy of target localization and the stability of classification judgment in vehicle-mounted infrared scenarios, and reducing the risk of false alarms and missed detections caused by uneven thermal radiation distribution and background interference.

[0015] Furthermore, step A5 also includes: The vectors with an exclusion index of 0 corresponding to the detection boxes in the classification confidence matrix are processed by the maximum value function to obtain the maximum class confidence score of the detection boxes; the detection boxes with the maximum class confidence score greater than the preset confidence threshold in the multi-class detection box parameter set are retained to obtain the high confidence detection result set; the high confidence detection result set is processed by non-maximum suppression operation to obtain the final detection result.

[0016] This invention also discloses an end-to-end infrared image super-resolution target detection system, comprising: Infrared image acquisition and coarse processing module: Acquires raw infrared images and performs preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain preprocessed high-resolution images and preprocessed raw resolution images; Artifact suppression module: Based on the preprocessed high-resolution image and the preprocessed original resolution image, the initial high-resolution depth feature map and the thermal radiation attention weighted map are extracted sequentially, and then the artifact suppression high-resolution feature map is obtained through the cross-scale thermal radiation consistency gating fusion mechanism. Multi-scale fusion module: Based on the artifact suppression high-resolution feature map and the preprocessed original resolution image, the module sequentially performs dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition, and cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. Detection module: Based on the fused multi-scale detection feature map, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set; Post-detection processing module: Based on the original detection result set, low-confidence detection boxes are removed sequentially, and overlapping detection boxes are removed by non-maximum suppression to obtain the final detection result.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) To address the artifact interference and false alarm / missed detection problems caused by the separation of image super-resolution reconstruction and target detection in traditional methods, this invention adopts an end-to-end infrared image super-resolution target detection architecture, which deeply integrates super-resolution reconstruction and target detection at the feature level, so that the super-resolution process is driven by the requirements of the detection task, rather than blindly pursuing pixel-level fidelity; in the super-resolution feature enhancement stage, by introducing a thermal radiation attention weighted map, real thermophysical information constraints are applied to the high-resolution feature enhancement process, so that the super-resolution network can distinguish between real target edges and reconstruction artifacts based on the statistical characteristics of thermal radiation, thereby suppressing the generation of false textures; at the same time, through dual-scale feature differentiation refinement and cross-level... Gated adaptive fusion complements the spatial details of high-resolution features with the fidelity of thermal radiation information of the original resolution features, enabling small targets at long distances to obtain sufficient spatial representation capabilities while avoiding the artifact risks of single high-resolution features. In the detection stage, a detection-specific tensor incorporating target spatial priors is constructed, and physical perception correction of bounding box regression and classification confidence is guided by thermal radiation adaptive modulation and thermal contrast, ensuring that detection decisions are consistent with infrared imaging patterns. Thus, this invention can effectively reduce false alarms caused by reconstruction artifacts in complex vehicle-mounted infrared scenes, while improving the detection stability of small targets under low contrast and partial occlusion conditions.

[0018] (2) In view of the problems of reconstruction artifacts interfering with detection decision and the removal of weak edge signals of small targets caused by separating super-resolution reconstruction and target detection in the prior art, the present invention designs the super-resolution reconstruction network to directly output detection-oriented high-resolution feature maps, so that the super-resolution process is driven by the requirements of the detection task; at the same time, it extracts thermal radiation attention weighted maps from the preprocessed original resolution images, encodes real thermal physical information based on local thermal radiation contrast maps and combined with local statistical enhancement feature maps, and fuses thermal radiation attention weighted maps with initial high-resolution depth feature maps through cross-scale thermal radiation consistency gating fusion mechanism, so that the high-resolution feature enhancement process is constrained by real thermal radiation physical information; thus, the super-resolution network can distinguish real target edges from reconstruction artifacts based on thermal radiation statistical characteristics, suppress the generation of false textures and artifacts, avoid the removal of weak edge signals of small targets, and generate feature expressions that are more favorable to detection.

[0019] (3) To address the problem that detection networks relying solely on single-scale feature inputs struggle to balance artifact risk and insufficient resolution, this invention designs a dual-stream complementary fusion architecture for high-resolution and original-resolution features. The high-resolution base feature map is refined using directional gradient response gating to enhance real structural edges and suppress false edges. The original-resolution base feature map is refined using local contrast normalization gating to amplify the difference between weak thermal signals and the surrounding background. Furthermore, the original-resolution hierarchical compensation generation network is guided by a high-resolution spatial attention map to inject spatial information, constraining the spatial expansion process of the original-resolution features by high-resolution edge information. Finally, cross-level gating adaptive fusion ensures that the advantages of the two scales complement each other and their disadvantages are mutually suppressed. Thus, distant small targets can obtain sufficient spatial representation capabilities while avoiding artifact risk in single-resolution features, reducing false alarms and missed detections caused by insufficient single-scale features.

[0020] (4) To address the problem that traditional detection heads are not adapted to infrared physical characteristics, resulting in unreliable positioning and classification confidence, this invention introduces thermal radiation priors from three levels: position encoding, bounding box regression, and classification confidence. By injecting thermal radiation differential position encoding into the target space prior, the detection head can perceive the thermal radiation intensity variation law corresponding to different spatial positions. A thermal radiation-guided YOLO bounding box regression head is used for bounding box regression, and the thermal radiation adaptive modulation coefficient map makes the bounding box regression focus more on the target area with real thermal radiation response. Through multi-level thermal contrast-guided classification confidence mapping, the classification decision is constrained by the thermal contrast physical information. Thus, the detection head can understand the physical law of thermal radiation changing with spatial position, the bounding box positioning can prioritize the response of areas with real thermal radiation characteristics, and the classification confidence can reflect the real thermal radiation difference between the target and the background, improving the accuracy of target positioning and the stability of classification judgment, and reducing the risk of false alarms and missed detections caused by uneven thermal radiation distribution. Attached Figure Description

[0021] Figure 1 A schematic flowchart of an end-to-end infrared image super-resolution target detection method provided by the present invention; Figure 2 The comparison between the original infrared image and the thermal radiation attention-weighted image provided by this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0023] Example 1: An end-to-end infrared super-resolution target detection method, such as... Figure 1 As shown, it includes the following steps: A1: Acquire raw infrared images and perform preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain preprocessed high-resolution images and preprocessed raw resolution images, including: A11: The original infrared image is acquired by the vehicle-mounted infrared detector, and then the noise is suppressed by the wavelet threshold denoising method. The thermal radiation intensity value is mapped to the preset value range by the linear normalization method to obtain the cleaned original resolution infrared image. A12: The original resolution infrared image after cleaning is spatially expanded using bicubic interpolation to generate an initial high-resolution image; then, the original resolution infrared image after cleaning is enhanced with histogram equalization to obtain an original resolution image with enhanced contrast. A13: By establishing a correspondence between the initial high-resolution image and the original resolution image with enhanced contrast through pixel-level spatial coordinate mapping, the alignment and association of multi-scale images are achieved, resulting in a preprocessed high-resolution image and a preprocessed original resolution image.

[0024] A2: Based on the preprocessed high-resolution image and the preprocessed original resolution image, an initial high-resolution depth feature map and a thermal radiation attention-weighted map are extracted sequentially. Then, an artifact-suppressed high-resolution feature map is obtained through a cross-scale thermal radiation consistency gating fusion mechanism, including: A21: Deep learning feature extraction is performed on the preprocessed high-resolution image to obtain an initial high-resolution depth feature map. The calculation method is as follows: ; ; ; ; in, This is a shallow convolutional feature map. This is a convolutional layer with a kernel size of 3×3. For the preprocessed high-resolution image, This is a deformation feature diagram. This is a deformable convolution operation. For GELU functions, For layer normalization, For channel attention gating weights, It is the Sigmoid activation function. This is a convolutional layer with a kernel size of 1×1. For global average pooling, Here is the initial high-resolution depth feature map, ⊕ represents element-wise addition, and ⊙ represents the Hadamard product; A22: Extract thermal physical information from the preprocessed original resolution image to obtain a thermal radiation attention-weighted map, calculated as follows: ; ; ; ; in, This is a local thermal radiation contrast map. To take the absolute value, This is the original resolution image after preprocessing. For a 3×3 window average pooling operation, For local statistical enhancement feature maps, This is a channel-level splicing operation. For a 3x3 window, max pooling operation. This is a feature map encoded by thermal radiation. For attention-weighted graphs of thermal radiation, such as Figure 2 As shown, the distant city buildings show a high response in the attention map - the boats on the water are significantly highlighted - the lower left shore area is also highlighted due to temperature differences - the low temperature water surface area is effectively suppressed (dark blue), making the network pay more attention to the heat source targets in the infrared image. A23: Based on the initial high-resolution depth feature map and the thermal radiation attention-weighted map, feature enhancement and super-resolution reconstruction are performed through a cross-scale thermal radiation consistency gated fusion mechanism to obtain an artifact-suppressed high-resolution feature map. The calculation method is as follows: ; ; ; ; in, This is the upsampled thermal radiation attention map. This is a bilinear upsampling operation. For cross-scale fusion of gated graphs To fuse and enhance feature maps, This is a high-resolution feature map for artifact suppression.

[0025] A3: Based on the artifact-suppressed high-resolution feature map and the preprocessed original resolution image, dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition, and cross-level gated adaptive fusion are performed sequentially to obtain a fused multi-scale detection feature map, including: A31: Based on the artifact-suppressed high-resolution feature map and the preprocessed original resolution image, a dual-stream initial feature extraction network is used. Then, directional gradient response gating and local contrast normalization gating are employed to differentially refine the dual-scale features, resulting in a high-resolution refined feature map and an original resolution refined feature map. The calculation method is as follows: ; ; ; ; in, For high-resolution basic feature maps, For convolutional downsampling networks, This is the original resolution base feature map. For high-resolution refined feature maps, It is a two-dimensional convolutional network. To calculate the gradient magnitude using the Sobel operator, LCN refines feature maps to the original resolution. Local contrast normalization; A32: Spatial information is injected based on the high-resolution refined feature map and the original resolution refined feature map to obtain the original resolution compensated feature map. The calculation method is as follows: ; ; ; in, For high-resolution spatial attention maps, This is the original resolution feature map after upsampling. This is a transposed convolution upsampling operation. The feature map is compensated for at the original resolution. A33: Spatial-channel dual-dimensional feature activation value competition is performed on the high-resolution refined feature map to obtain a high-resolution salient response feature map. The calculation method is as follows: ; ; in, For high-resolution spatially significant response distribution, For group normalization operation, This is the temperature scaling factor. High-resolution salient response feature map; A34: The high-resolution salient response feature map and the original resolution-compensated feature map are fused through cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. The calculation method is as follows: ; ; in, For cross-level gated adaptive fusion vectors, To fuse multi-scale detection feature maps.

[0026] Specifically, for scenarios where there are significant differences in local thermal radiation consistency between the high-resolution significant response feature map and the original resolution compensated feature map, this invention also provides a cross-level fusion calculation method based on thermal radiation local consistency gating to replace the A34 step. The calculation method is as follows: ; ; ; in, This is a diagram showing the differences in thermal radiation characteristics across different levels. This is a gating system for localized consistency of thermal radiation across different levels.

[0027] A4: Based on the fused multi-scale detection feature maps, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set, including: A41: Based on the fused multi-scale detection feature map, a priori object space is injected through thermal radiation differential position encoding, and then a detection-specific tensor is constructed by combining it with dimension alignment mapping. The calculation method is as follows: ; ; ; in, For location coding based on thermal radiation differences, It is a sine function. A function for generating coordinates for a two-dimensional grid. To fuse the height of multi-scale detection feature maps, To fuse the widths of multi-scale detection feature maps, This is the thermal radiation temperature scaling factor. The intermediate features are after coordinate enhancement. To detect special tensors, To reduce the placement of the confidence coefficient; A42: Based on the detection-specific tensor, a thermally guided YOLO bounding box regression head is used to perform bounding box regression, obtaining a multi-class detection box parameter set. The calculation method is as follows: ; ; ; in, This is a diagram of the adaptive modulation coefficients for thermal radiation. The intermediate features of the YOLO regression branch after thermal radiation weighting are... For multi-class detection box parameter set, This is the scaling factor for the bounding box parameters; A43: Based on the multi-class detection box parameter set, the thermally deformed corrected detection features are fused with the detection-specific tensor through multi-level thermal contrast-guided classification confidence mapping. The classification confidence is calculated and integrated into the original detection result set. The calculation method is as follows: ; ; ; in, This is a diagram showing the thermal contrast guiding coefficient. For the classification confidence matrix, For the Softmax function, This is the set of original detection results. This is the maximum index operation.

[0028] A5: Based on the original detection result set, low-confidence detection boxes are removed sequentially, and overlapping detection boxes are then removed using non-maximum suppression to obtain the final detection results, including: A51: Based on the original detection result set, calculate the confidence level of the classification prediction vector for each detection box, and remove low-confidence detection boxes to obtain the high-confidence detection result set. The calculation method is as follows: ; ; in, Let be the maximum class confidence score of detection box b. This is the vector in the classification confidence matrix corresponding to the exclusion index 0 of the detection box b. To find the maximum value function, This is the set of high-confidence detection results, where θ is the preset confidence threshold. A52: Based on the high-confidence detection result set, overlapping detection boxes are removed using non-maximum suppression to obtain the final detection result. The calculation method is as follows: ; in, For the final test results, This is a nonmaximum suppression operation.

[0029] In a preferred embodiment of the present invention, the parameters of the above-mentioned end-to-end infrared image super-resolution target detection network are configured and trained, and the specific settings of each module are as follows: In the preprocessing stage, the upsampling factor for spatial resolution augmentation using bicubic interpolation is set to 2x, and the preset numerical range for linear normalization is set to 0 to 1. In the super-resolution feature enhancement stage, the preprocessed high-resolution image is input to a 3×3 convolutional layer with 64 output channels. The output channels of the deformable convolution operation are also set to 64, and the output channels of the 3×3 convolutional layer in the offset generation network are set to 18. After global average pooling, a 1×1 convolutional layer is applied, with the output channels first reduced to 16 and then expanded to 64. The output channels of the channel attention gating weights are set to 64. The output channels of the initial high-resolution depth feature map are also set to 64. The number of channels is set to 64; the preprocessed original resolution image is concatenated by the channel dimensions of the local thermal radiation contrast map, the average pooling operation result of the 3×3 window, and the max pooling operation result of the 3×3 window, and then input into a convolutional layer with a kernel size of 3×3, with the output channel number set to 64; the output channel number of the thermal radiation encoded feature map is set to 64; the output channel number of the thermal radiation attention weighted map is set to 64; the upsampling factor of the bilinear upsampling operation is set to 2 times; the output channel number of the cross-scale fusion gating map is set to 64; the output channel number of the artifact suppression high-resolution feature map is set to 64; in the multi-scale feature fusion stage, the convolutional downsampling network contains two convolutional layers with a kernel size of 3×3, each... The convolutional layer is followed by a downsampling operation with a stride of 2, resulting in a total downsampling factor of 4, and the number of output channels for each convolutional layer is set to 256; the number of output channels for the 2D convolutional network in the directional gradient response gating is set to 256; the local contrast normalization gating is followed by a convolutional layer with a kernel size of 3×3, and the number of output channels is set to 256; the number of output channels for the high-resolution spatial attention map is set to 1; the number of output channels for the transposed convolution upsampling operation is set to 256, and the upsampling factor is set to 2; the number of output channels for the high-resolution spatial salient response distribution is set to 1; the number of groups for the group normalization operation is set to 32; the temperature scaling factor is set to 0.1; and the number of output channels for the cross-layer gated adaptive fusion vector is set to 1. The number of output channels for the fused multi-scale detection feature map is set to 256; during the detection phase, the number of output channels for the 1×1 convolutional layer in the thermal radiation differential location encoding is set to 256; the thermal radiation temperature scaling factor is set to 10000; the shrinkage confidence coefficient is initialized to 1.0 and updated as a learnable parameter during training; the number of output channels for the detection-specific tensor is set to 256; the number of output channels for the thermal radiation adaptive modulation coefficient map is set to 256; the number of output channels for the intermediate features of the thermal radiation-weighted YOLO regression branch is set to 256; the number of output channels for the multi-class detection box parameter set is set to 4; and the bounding box parameter scaling factors are initialized to 0.5, 0.5, 1.0, and 1.0; The number of output channels for the thermal contrast guided coefficient map is set to 1; The number of output channels for the classification confidence matrix is ​​set to the number of categories of the target to be detected; In the post-processing stage, the preset confidence threshold is set to 0.5; The threshold for non-maximum suppression operation is set to 0.5; In the training stage, the entire network is jointly trained end-to-end using the Adam optimizer, with an initial learning rate of 0. The learning rate is reduced to 0.1 times its original value every 50 training rounds. The batch size is set to 8, the total number of training rounds is set to 200, and the loss function includes bounding box regression loss and classification loss. The two losses are weighted and summed according to preset weights.

[0030] Example 2: The present invention also discloses an end-to-end infrared image super-resolution target detection system, comprising: Infrared image acquisition and coarse processing module: Acquires raw infrared images and performs preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain preprocessed high-resolution images and preprocessed raw resolution images; Artifact suppression module: Based on the preprocessed high-resolution image and the preprocessed original resolution image, the initial high-resolution depth feature map and the thermal radiation attention weighted map are extracted sequentially, and then the artifact suppression high-resolution feature map is obtained through the cross-scale thermal radiation consistency gating fusion mechanism. Multi-scale fusion module: Based on the artifact suppression high-resolution feature map and the preprocessed original resolution image, the module sequentially performs dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition, and cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. Detection module: Based on the fused multi-scale detection feature map, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set; Post-detection processing module: Based on the original detection result set, low-confidence detection boxes are removed sequentially, and overlapping detection boxes are removed by non-maximum suppression to obtain the final detection result.

[0031] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0032] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0033] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An end-to-end infrared image super-resolution target detection method, characterized in that, Includes the following steps: A1: Acquire the original infrared image and perform preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain the preprocessed high-resolution image and the preprocessed original resolution image; A2: Based on the preprocessed high-resolution image and the preprocessed original resolution image, the initial high-resolution depth feature map and the thermal radiation attention weighted map are extracted sequentially, and then the artifact-suppressed high-resolution feature map is obtained through the cross-scale thermal radiation consistency gating fusion mechanism. A3: Based on the artifact suppression high-resolution feature map and the preprocessed original resolution image, dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition and cross-level gating adaptive fusion are performed sequentially to obtain a fused multi-scale detection feature map. A4: Based on the fused multi-scale detection feature map, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set; A5: Based on the original detection result set, low-confidence detection boxes are removed sequentially, and overlapping detection boxes are removed by non-maximum suppression to obtain the final detection result.

2. The end-to-end infrared image super-resolution target detection method according to claim 1, characterized in that, Step A1 includes: A11: The original infrared image is acquired by the vehicle-mounted infrared detector, and then the noise is suppressed by the wavelet threshold denoising method. The thermal radiation intensity value is mapped to the preset value range by the linear normalization method to obtain the cleaned original resolution infrared image. A12: The original resolution infrared image after cleaning is spatially expanded using bicubic interpolation to generate an initial high-resolution image; then, the original resolution infrared image after cleaning is enhanced with histogram equalization to obtain an original resolution image with enhanced contrast. A13: By establishing a correspondence between the initial high-resolution image and the original resolution image with enhanced contrast through pixel-level spatial coordinate mapping, the alignment and association of multi-scale images are achieved, resulting in a preprocessed high-resolution image and a preprocessed original resolution image.

3. The end-to-end infrared image super-resolution target detection method according to claim 1, characterized in that, Step A2 includes: A21: Deep learning feature extraction is performed on the preprocessed high-resolution image to obtain an initial high-resolution depth feature map. The calculation method is as follows: ； ； ； ； in, This is a shallow convolutional feature map. This is a convolutional layer with a kernel size of 3×3. For the preprocessed high-resolution image, This is a deformation feature diagram. This is a deformable convolution operation. For GELU functions, For layer normalization, For channel attention gating weights, It is the Sigmoid activation function. This is a convolutional layer with a kernel size of 1×1. For global average pooling, Here is the initial high-resolution depth feature map, ⊕ represents element-wise addition, and ⊙ represents the Hadamard product; A22: Extract thermal physical information from the preprocessed original resolution image to obtain a thermal radiation attention-weighted map, calculated as follows: ； ； ； ； in, This is a local thermal radiation contrast map. To take the absolute value, This is the original resolution image after preprocessing. For a 3×3 window average pooling operation, For local statistical enhancement feature maps, This is a channel-level splicing operation. For a 3x3 window, max pooling operation. This is a feature map encoded by thermal radiation. Attention-weighted plot of thermal radiation; A23: Based on the initial high-resolution depth feature map and the thermal radiation attention-weighted map, feature enhancement and super-resolution reconstruction are performed through a cross-scale thermal radiation consistency gated fusion mechanism to obtain an artifact-suppressed high-resolution feature map. The calculation method is as follows: ； ； ； ； in, This is the upsampled thermal radiation attention map. This is a bilinear upsampling operation. For cross-scale fusion of gated graphs To fuse and enhance feature maps, This is a high-resolution feature map for artifact suppression.

4. The end-to-end infrared image super-resolution target detection method according to claim 3, characterized in that, Step A3 includes: A31: Based on the artifact-suppressed high-resolution feature map and the preprocessed original resolution image, a dual-stream initial feature extraction network is used. Then, directional gradient response gating and local contrast normalization gating are employed to differentially refine the dual-scale features, resulting in a high-resolution refined feature map and an original resolution refined feature map. The calculation method is as follows: ； ； ； ； in, For high-resolution basic feature maps, For convolutional downsampling networks, This is the original resolution base feature map. For high-resolution refined feature maps, It is a two-dimensional convolutional network. To calculate the gradient magnitude using the Sobel operator, LCN refines feature maps to the original resolution. Local contrast normalization; A32: Spatial information is injected based on the high-resolution refined feature map and the original resolution refined feature map to obtain the original resolution compensated feature map. The calculation method is as follows: ； ； ； in, For high-resolution spatial attention maps, This is the original resolution feature map after upsampling. This is a transposed convolution upsampling operation. The feature map is compensated for at the original resolution. A33: Spatial-channel dual-dimensional feature activation value competition is performed on the high-resolution refined feature map to obtain a high-resolution salient response feature map. The calculation method is as follows: ； ； in, For high-resolution spatially significant response distribution, For group normalization operation, This is the temperature scaling factor. High-resolution salient response feature map; A34: The high-resolution salient response feature map and the original resolution-compensated feature map are fused through cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. The calculation method is as follows: ； ； in, For cross-level gated adaptive fusion vectors, To fuse multi-scale detection feature maps.

5. The end-to-end infrared image super-resolution target detection method according to claim 4, characterized in that, The A4 step includes: A41: Based on the fusion of multi-scale detection feature maps, target space priors are injected through thermal radiation differential position encoding, and then a detection-specific tensor is constructed by combining dimension alignment mapping. A42: Based on the detection-specific tensor, the thermal radiation-guided YOLO bounding box regression head is used to perform bounding box regression, resulting in a multi-class detection box parameter set; A43: Based on the multi-class detection box parameter set, the thermal deformation corrected detection features are fused with the detection-specific tensor through multi-level thermal contrast-guided classification confidence mapping, the classification confidence is calculated and integrated into the original detection result set.

6. The end-to-end infrared image super-resolution target detection method according to claim 5, characterized in that, The A4 step includes: The height and width of the fused multi-scale detection feature map are concatenated, processed by a 2D grid coordinate generation function, then processed by a 1×1 convolutional layer, divided by a thermal radiation temperature scaling factor, and then processed by a sine function to obtain the thermal radiation differential position code. The fused multi-scale detection feature map is then processed by a 1×1 convolutional layer, and added element-wise with the Hadamard product of the reduced confidence coefficient and the thermal radiation differential position code to obtain the coordinate-enhanced intermediate feature. The coordinate-enhanced intermediate feature is then added element-wise with the fused multi-scale detection feature map, and then processed by layer normalization to obtain the detection-specific tensor. The detection-specific tensor is processed by a 1×1 convolutional layer and then by a Sigmoid activation function to obtain a thermally adaptive modulation coefficient map. The detection-specific tensor and the thermally adaptive modulation coefficient map are then multiplied by a Hadamard layer and processed by a 3×3 convolutional layer to obtain thermally weighted YOLO regression branch intermediate features. The thermally weighted YOLO regression branch intermediate features are then multiplied by a 1×1 convolutional layer and then multiplied by a bounding box parameter scaling factor to obtain a multi-class detection box parameter set. The detection-specific tensor is element-wise summed with the intermediate features of the thermally weighted YOLO regression branch, then processed by a 1×1 convolutional layer, followed by a Sigmoid activation function to obtain the thermal contrast guiding coefficient map. The intermediate features of the thermally weighted YOLO regression branch are processed by a 1×1 convolutional layer, then combined with the thermal contrast guiding coefficient map using a Hadamard product, followed by a Softmax function to obtain the classification confidence matrix. The multi-class detection box parameter set, the classification confidence matrix, and the result of the maximum indexing operation on the classification confidence matrix are integrated to obtain the original detection result set.

7. The end-to-end infrared image super-resolution target detection method according to claim 6, characterized in that, Step A5 includes: The vectors with an exclusion index of 0 corresponding to the detection boxes in the classification confidence matrix are processed by the maximum value function to obtain the maximum class confidence score of the detection boxes; the detection boxes with the maximum class confidence score greater than the preset confidence threshold in the multi-class detection box parameter set are retained to obtain the high confidence detection result set; the high confidence detection result set is processed by non-maximum suppression operation to obtain the final detection result.

8. An end-to-end infrared super-resolution target detection system, characterized in that, include: Infrared image acquisition and coarse processing module: Acquires raw infrared images and performs preprocessing, spatial resolution augmentation, local contrast enhancement, and multi-scale image alignment and correlation to obtain preprocessed high-resolution images and preprocessed raw resolution images; Artifact suppression module: Based on the preprocessed high-resolution image and the preprocessed original resolution image, the initial high-resolution depth feature map and the thermal radiation attention weighted map are extracted sequentially, and then the artifact suppression high-resolution feature map is obtained through the cross-scale thermal radiation consistency gating fusion mechanism. Multi-scale fusion module: Based on the artifact suppression high-resolution feature map and the preprocessed original resolution image, the module sequentially performs dual-scale feature differentiation refinement, spatial information injection, spatial-channel dual-dimensional feature activation value competition, and cross-level gated adaptive fusion to obtain a fused multi-scale detection feature map. Detection module: Based on the fused multi-scale detection feature map, a detection-specific tensor is constructed sequentially, bounding box regression is performed to obtain a multi-class detection box parameter set, and then the classification confidence is calculated and integrated into the original detection result set; Post-detection processing module: Based on the original detection result set, low-confidence detection boxes are sequentially removed, and overlapping detection boxes are removed by non-maximum suppression to obtain the final detection result; so as to realize the end-to-end infrared image super-resolution target detection method as described in any one of claims 1-7.

Images