A low-contrast target detection method based on a shrinkage guided core point set representation

Abstract

The application discloses a low-contrast target detection method based on a shrinkage guide core point set representation, and relates to the technical fields of computer vision and target detection.The application introduces a shrinkage prior box guided point set generation mechanism, a point set optimization strategy based on Gaussian distribution consistency, and a minimum circumscribed rectangle regression method, realizes accurate description and stable positioning of the spatial structure of a target in a low-contrast scene, and can effectively overcome the inherent defects of the poor geometric adaptability of a traditional rectangular box representation under the conditions of boundary blur and texture loss, solves key problems such as feature flooding, positioning drift and high missing detection rate in a low signal-to-noise ratio environment from the aspect of target representation, and systematically solves the key problems. The method can be used for visual tasks under complex light conditions such as automatic driving, intelligent security, remote sensing image analysis and industrial visual detection, and significantly improves the positioning accuracy, classification confidence and cross-scene generalization ability of a detection model in a weak contrast environment.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to a low-contrast target detection method based on shrinking guided core point set representation, which is applicable to application scenarios requiring high-precision target positioning, such as autonomous driving, video surveillance, and remote sensing image analysis. Background Technology

[0002] In recent years, object detection, as a core task in the field of computer vision, has been widely applied in scenarios such as autonomous driving, video surveillance, and remote sensing analysis. Existing mainstream methods are based on anchor boxes or keypoint representations, which perform well in normal lighting and high-contrast scenes. However, under low-contrast conditions, such as insufficient image illumination and weak differences between the target and the background, the edge and texture information of the target are severely degraded, making it difficult for traditional methods to accurately locate the target boundary, leading to frequent problems such as missed detections, localization errors, and classification mistakes.

[0003] To address the challenge of low contrast, existing research has attempted to improve detection performance through image preprocessing enhancement or feature adaptation mechanisms. However, most of these methods still rely on traditional bounding box representations, failing to address the boundary ambiguity problem at the geometric representation level. Bounding box representations are prone to introducing background noise or losing target information when the target shape is irregular or the boundaries are unclear. Especially in low-contrast scenes, their localization accuracy is severely limited, making it difficult to meet practical application requirements.

[0004] Therefore, there is an urgent need for a target representation method that can adapt to low-contrast scenes and has stronger geometric adaptability. Point set representation can flexibly characterize the shape of the target, and combined with prior geometric constraints and distribution optimization, it is expected to fundamentally improve the robustness of the model under complex lighting conditions. However, how to achieve stable generation and fine-tuning of point sets in low signal-to-noise ratio environments remains a key technical challenge. Summary of the Invention

[0005] The purpose of this invention is to propose a low-contrast target detection method based on shrinking guided core point set representation to solve the problems mentioned in the background art. By representing the target as a learnable core prior quadrilateral, and combining geometric prior and distribution consistency constraints, the method achieves accurate characterization and stable regression of the target's spatial structure under low contrast, effectively improving detection accuracy and generalization ability, and providing innovative technical support for visual perception systems under complex lighting conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A low-contrast target detection method based on shrinking guided core point set representation specifically includes the following steps: S1. Dataset Construction: Based on publicly available low-light target detection benchmark datasets, the datasets are preprocessed to obtain low-light images to be detected. S2. Network Structure Construction: A single-stage target detection network model for low-contrast scenes is proposed, which follows the classic Backbone-FPN-Head paradigm. The detection head of the network model adopts a target regression method based on shrinking guided core point set representation to replace the traditional anchor box regression or key point estimation mechanism. S3. Loss Function Design: Construct a multi-objective joint loss function system to jointly constrain the generated results at the geometric structure and semantic alignment levels, so as to achieve synergistic optimization of point set generation, distribution constraints and detection accuracy; S4. Model Training and Validation: Based on S1~S3, a complete model training and validation process is designed. In the training phase, the model is input with low-light images and their corresponding ground truth bounding boxes and class labels in batches. The prediction results and various losses are calculated through forward propagation, and the network parameters are updated using the backpropagation algorithm. Multi-scale training strategies and data augmentation methods are adopted to enhance the adaptability and generalization of the model. In the validation phase, the model relies on the input images to perform a single forward inference and outputs the coordinates of the detection boxes and the class confidence.

[0008] Preferably, the dataset preprocessing in S1 specifically includes: All images in the dataset are uniformly resized to a fixed input size and then normalized using mean-variance. A random partitioning strategy is adopted, with 80% of the images used for training and 20% for testing, to ensure a balanced distribution of samples for each category under different lighting conditions, providing a stable and reliable data foundation for model training and performance evaluation.

[0009] Preferably, the single-stage object detection network model for low-contrast scenes described in S2 specifically includes: Backbone Network and Feature Pyramid Module: A pre-trained ResNet-50 is used as the backbone network to extract multi-scale deep features from the input image, and multi-scale feature maps are generated through the Feature Pyramid Network (FPN) to provide rich semantic and spatial information for subsequent point set regression. The shrinking guided core point set generation module based on shrinking prior boxes is used to receive the feature map output by the Feature Pyramid Network (FPN), output the initial shrinking guided core point coordinates through a lightweight convolutional prediction head, and introduce a prior box obtained by shrinking the ground truth labeled box as a strong geometric constraint. By fusing the initial predicted point set coordinates with the shrinking prior box, the four points of the final shrinking guided core point set are strictly constrained to the straight lines of the four boundaries of the shrinking prior box. Connecting the four points forms a core prior quadrilateral to enhance the feature focusing ability of the target core region. The core prior quadrilateral optimization module based on Gaussian distribution consistency models the core prior quadrilateral and the shrinking prior box as two-dimensional Gaussian distributions respectively, and achieves fine-tuning of the point set by minimizing the distribution difference between the two. The target regression module based on the minimum bounding rectangle of the shrinking guide core point set: The minimum bounding rectangle method is used to construct the prediction box by taking the minimum and maximum horizontal and vertical coordinates of the shrinking guide core point set, so as to achieve joint optimization from the generation of the shrinking guide core point set to the output of the final detection box.

[0010] Preferably, the multi-objective joint loss function system includes classification loss and overall regression loss, and the comprehensive constraint is achieved through a weighted summation. The classification loss adopts Focal Loss to alleviate the problem of extreme imbalance between positive and negative samples in low-contrast scenes, so that the model pays more attention to the difficult-to-classify samples and improves the classification confidence and discrimination ability. The overall regression loss is composed of a weighted sum of the bounding box regression loss and the Gaussian distribution similarity loss, with the weights determined through cross-validation tuning.

[0011] Preferably, the bounding box regression loss uses Smooth L1 loss to constrain the coordinate difference between the predicted box and the ground truth box, so as to ensure the localization accuracy of the detected box and reduce the impact of outliers on gradient updates. The Gaussian distribution similarity loss, based on Hellinger distance, is used to constrain the contraction of the core prior quadrilateral and guide the core point set distribution to converge toward the contracted prior box, thereby enhancing the spatial stability and geometric compactness of the point set under low contrast.

[0012] Preferably, during the training phase, a stochastic gradient descent optimizer is used, with momentum set to 0.9 and weight decay to 0.0001; the initial learning rate is set to 0.001, a cosine annealing scheduling strategy is adopted, and the learning rate decreases by a factor of 0.1 in the 18th and 22nd rounds; the total number of training rounds is 100, and the batch size is 8.

[0013] Preferably, in the verification phase, the average accuracy is used as the evaluation metric for the detection performance of the model in low-contrast scenarios.

[0014] The present invention further protects a computer device, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the instruction, program, code set or instruction set being loaded and executed by the processor to implement the above-mentioned low-contrast target detection method based on shrinking guide core point set representation.

[0015] The present invention further protects a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the instruction, program, code set, or instruction set is loaded and executed by a processor to implement the above-described low-contrast target detection method based on a contracted guide core point set representation.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a low-contrast target detection method based on a contracted guided core point set representation, achieving a fundamental breakthrough over the traditional rectangular box paradigm in target geometric representation. The contracted guided core point set representation proposed in this invention possesses stronger shape adaptability and boundary robustness, and can stably characterize the target's spatial structure even under conditions of blurred edges, missing textures, and extremely low signal-to-noise ratios, fundamentally improving the detection model's positioning accuracy and anti-interference capability in complex lighting environments. Specifically, it has the following significant beneficial effects: (1) Significantly improves detection accuracy in low-contrast scenes. On the ExDark public dataset, this invention leads all comparison methods with an mAP of 71.1%@0.5, especially achieving breakthrough improvements in traditionally difficult detection categories such as Bottle, Table, and People, with significant reductions in false negative rate and localization error.

[0017] (2) A complete "geometric prior + distribution optimization" shrinkage-guided core point set regression technology system was constructed. The shrinkage prior box guidance mechanism provides clear geometric anchors for feature-blurred regions under low contrast, fundamentally suppressing point set divergence; Gaussian distribution consistency optimization constrains the shrinkage-guided core point set to the target core region through distribution similarity measurement, achieving adaptive fitting and stable convergence of the target shape. The two work together to enable the model to output compact, accurate, and semantically consistent detection boxes even under extreme degradation conditions.

[0018] (3) It expands the new paradigm of target geometric representation. This invention introduces the "shrinking guided core point set + Gaussian distribution" mechanism into the low-contrast target detection task for the first time, providing a new technical path for subsequent research. This method is not only applicable to low-light environments, but can also be extended to generalized low-contrast visual tasks such as haze imaging, underwater detection, infrared imaging, and medical imaging, and has broad cross-domain application prospects.

[0019] In summary, this invention systematically solves the core bottlenecks of insufficient target representation capability and poor localization robustness in low-contrast scenes by constructing a target geometric modeling and optimization framework based on shrinking guided core point set representation. It provides a high-precision and highly robust innovative technical solution for the engineering implementation of visual perception systems in complex environments. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of a low-contrast target detection method based on the representation of a contracted guided core point set proposed in this invention. Figure 2 This is a schematic diagram of the core prior quadrilateral target representation region proposed in Embodiment 2 of the present invention; Figure 3 This is a visualization of the detection results on the ExDark dataset proposed in Embodiment 2 of the present invention. Detailed Implementation

[0021] 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.

[0022] This invention proposes a low-contrast target detection method based on contraction-guided core point set representation. By introducing a point set generation mechanism guided by contraction prior boxes, a point set optimization strategy based on Gaussian distribution consistency, and a minimum bounding rectangle regression method, it achieves accurate characterization and stable localization of the target's spatial structure in low-contrast scenes. The proposed method effectively overcomes the inherent defects of traditional rectangular box representation in terms of insufficient geometric adaptability under conditions of blurred boundaries and missing textures. Starting from the target representation level, it systematically solves key problems such as feature submersion, localization drift, and high false negative rates in low signal-to-noise ratio environments. This method can be applied to visual tasks under complex lighting conditions, such as autonomous driving, intelligent security, remote sensing image analysis, and industrial visual inspection. It significantly improves the localization accuracy, classification confidence, and cross-scene generalization ability of the detection model in low-contrast environments, thus providing highly robust technical support for the engineering implementation of low-quality visual perception systems. The specific details are as follows.

[0023] Example 1: This invention proposes a low-contrast target detection method based on a contracted guided core point set representation, specifically including the following: 1. Introduction to the dataset The dataset used in this invention is the publicly available low-light object detection benchmark dataset ExDark, used to verify the applicability of the model in real-world low-contrast environments. The ExDark dataset contains 7,363 images, covering 10 different lighting conditions from extremely dark environments to dusk. The dataset is labeled with 12 common object categories, including bicycles, boats, bottles, buses, cars, cats, chairs, cups, dogs, motorcycles, people, and tables. These categories exhibit varying degrees of appearance degradation, edge blurring, and texture loss under low-light conditions, providing a comprehensive challenge for verifying the model's robustness in low-contrast scenes.

[0024] During data preprocessing, all images were uniformly resized to a fixed input size and normalized using mean-variance. To preserve the true degradation characteristics of the original low-contrast scenes, no additional artificial enhancement or denoising operations were performed. The dataset employed a random partitioning strategy, with 80% of the images used for training and 20% for testing. This ensured a balanced distribution of samples across categories under different lighting conditions, providing a stable and reliable data foundation for model training and performance evaluation.

[0025] 2. Build the network structure This invention proposes a single-stage object detection network architecture for low-contrast scenes, generally following the classic Backbone-FPN-Head paradigm. Its core innovation lies in employing an object regression method based on a contracted guided core point set representation in the detection head, replacing traditional anchor box regression or keypoint estimation mechanisms. The network structure mainly consists of the following modules: (1) Backbone network and feature pyramid A pre-trained ResNet-50 is used as the backbone network to extract multi-scale deep features from the input image, and multi-scale feature maps are generated through the Feature Pyramid Network (FPN) to provide rich semantic and spatial information for subsequent point set regression.

[0026] (2) Generation of core point set of contraction guidance based on contraction prior box This module receives the feature map output by the FPN and outputs the coordinates of the initial shrinkage guiding core points through a lightweight convolutional prediction head. To overcome the problem of point set divergence caused by feature blurring and edge loss under low contrast, this invention introduces a prior box obtained by shrinking the ground truth bounding box as a strong geometric constraint. By fusing the coordinates of the initial predicted point set with the shrinkage prior box, the four points of the final shrinkage guiding core point set are strictly constrained to the straight lines containing the four boundaries of the shrinkage prior box. Connecting the four points forms a core prior quadrilateral, thereby enhancing the feature focusing ability of the target core region.

[0027] (3) Core Prior Quadrilateral Optimization Based on Gaussian Distribution Consistency To further improve the distribution stability and semantic fidelity of the shrinking guide core point set in low-contrast scenes, this invention models the core prior quadrilateral and the shrinking prior box as two-dimensional Gaussian distributions, and achieves fine-tuning of the point set by minimizing the distribution difference between them. This optimization mechanism not only preserves the point set's adaptive expressive ability to the target shape, but also significantly enhances the spatial stability and noise resistance of the point set under low contrast conditions.

[0028] (4) Target regression based on the minimum bounding rectangle of the shrinking core point set The optimized core prior quadrilateral needs to be converted into a standard axis aligned detection box to participate in subsequent non-maximum suppression and loss calculation. This invention uses the minimum bounding rectangle method, taking the minimum and maximum x and y coordinates of the contracted guiding core point set to form the prediction box. This conversion process is fully differentiable and supports end-to-end gradient backpropagation, thereby achieving joint optimization from the generation of the contracted guiding core point set to the final detection box output.

[0029] 3. Loss Function Design To achieve synergistic optimization of point set generation, distribution constraints, and detection accuracy, this invention constructs a multi-objective joint loss function system to jointly constrain the generated results at the geometric structure and semantic alignment levels. The overall loss function consists of multiple sub-modules, which achieve comprehensive constraints through a weighted summation, mainly including: (1) Classification loss: Focal Loss is adopted to alleviate the problem of extreme imbalance between positive and negative samples in low contrast scenes, so that the model pays more attention to difficult-to-classify samples and improves classification confidence and discrimination ability.

[0030] (2) Overall regression loss: This is a weighted sum of the bounding box regression loss and the Gaussian distribution similarity loss, with the weights determined through cross-validation tuning. The bounding box regression loss uses the Smooth L1 loss to constrain the coordinate differences between the predicted and ground truth boxes, ensuring the localization accuracy of the detected boxes while reducing the impact of outliers on gradient updates. The Gaussian distribution similarity loss, based on Hellinger distance, constrains the contraction of the core prior quadrilateral, guiding the core point set distribution to converge towards the contracted prior box, thus enhancing the spatial stability and geometric compactness of the point set under low contrast. The two losses are jointly optimized through weighted optimization to improve the localization quality of the detected boxes and the robustness of the contracted core point set representation.

[0031] The weight coefficients of each loss term are determined through cross-validation tuning and jointly optimized end-to-end during training to ensure that the model achieves the best balance between classification confidence, localization accuracy and geometric stability.

[0032] 4. Model Training and Validation This invention designs a complete model training and verification process to ensure that the proposed target detection method based on shrinkage-guided core point set representation has the best performance and the strongest generalization ability in low-contrast scenes.

[0033] (1) Training phase During training, the model was fed low-light images along with their corresponding ground truth bounding boxes and class labels in batches. Forward propagation was used to calculate predictions and various losses, while backpropagation was used to update network parameters. A stochastic gradient descent optimizer was used with a momentum of 0.9 and a weight decay of 0.0001. The initial learning rate was set to 0.001, and a cosine annealing scheduling strategy was employed, decreasing by a factor of 0.1 at epochs 18 and 22. The total training epochs were 100, with a batch size of 8. All experiments were implemented using the PyTorch deep learning framework and conducted in a distributed manner on an NVIDIA GeForce RTX 3090 GPU.

[0034] To enhance the model's adaptability to low-contrast scenes, a multi-scale training strategy was employed during training, with the shorter side of the input image randomly sampled from the interval [480, 960]. Simultaneously, data augmentation techniques such as random flipping and color dithering were used to improve the model's generalization ability without compromising the original low-contrast degradation characteristics. During training, the trends of classification loss, regression loss, and total loss curves were recorded to monitor the model's convergence status and gradient stability in real time.

[0035] (2) Verification phase During the validation phase, the model performs a single forward inference based solely on the input image, outputting the bounding box coordinates and class confidence score. To comprehensively evaluate the model's detection performance in low-contrast scenes, this invention employs a common evaluation metric in the object detection field: mean accuracy (mAP@0.5).

[0036] (3) Experimental results Quantitative experimental results on the ExDark low-light dataset show that the target detection method based on the shrinking guided core point set representation proposed in this invention achieves 71.1% mAP@0.5 on the ExDark low-light dataset, which is significantly better than the mainstream comparison methods. In particular, it has made significant improvements in categories that are easy to miss and difficult to locate under low contrast, such as Bottle, Table, and People. This verifies the superiority of the shrinking guided core point set representation and Gaussian optimization mechanism under extreme degradation conditions.

[0037] Example 2: Based on Embodiment 1, but with some differences, the method proposed in this invention will be further described below with reference to the relevant accompanying drawings and specific examples, specifically including the following: 1. Introduction to the dataset This invention uses the publicly available low-light object detection benchmark dataset ExDark for model training and performance validation. The ExDark dataset is specifically designed for low-contrast computer vision tasks and contains 7,363 images covering 10 different lighting conditions from extremely dark environments to dusk, including low light, ambient light, object light source, single light source, weak light, strong light, screen light, window light, shadow, and twilight. The dataset is labeled with 12 common object categories: bicycle, boat, bottle, bus, car, cat, chair, cup, dog, motorcycle, person, and table. These categories exhibit varying degrees of appearance degradation, edge blurring, and texture loss under low-light conditions, effectively testing the model's robustness and generalization ability in low-contrast scenes.

[0038] During data preprocessing, all images were uniformly adjusted to a shorter side of 640 pixels and subjected to mean-variance standardization. To preserve the realistic degradation characteristics of the original low-contrast scene, no additional artificial enhancement, denoising, or artifact suppression operations were performed. The dataset was randomly divided into 80% training and 20% testing portions to ensure a balanced distribution of samples for each category under different lighting conditions, providing a stable and reliable data foundation for model training and performance evaluation.

[0039] 2. Network Structure Setup This invention proposes a single-stage object detection network architecture for low-contrast scenes, following the Backbone-FPN-Head paradigm. Its core innovation lies in employing an object regression method based on a contracted guided core point set representation in the detection head, replacing traditional anchor box regression or keypoint estimation mechanisms. The network structure is as follows: Figure 1 As shown, it mainly consists of the following modules: (1) Backbone network and feature pyramid A ResNet-50 pre-trained on ImageNet is used as the backbone network to extract multi-scale depth features from the input image. The backbone network outputs feature maps at layers C3, C4, and C5, which are then processed by a Feature Pyramid Network (FPN) to generate multi-scale features at layers P3 to P7. The downsampling strides of each layer relative to the input image are 8, 16, 32, 64, and 128, respectively. The number of channels in each feature map is uniformly 256. The shrinkage-guided core point set regression module proposed in this invention is deployed independently in each layer of the FPN detection head to achieve joint optimization of multi-scale targets.

[0040] (2) Generation of core point set of contraction guidance based on contraction prior box To accurately characterize the spatial structure of targets in low-contrast scenes, this invention proposes a core prior quadrilateral generation method guided by a shrinking prior box. This method uses the shrinking prior box as a geometric constraint, binding the point set generation to the straight line of the shrinking prior box boundary, thereby enhancing the feature focusing capability of the target's core region.

[0041] The feature map input is received through a lightweight convolutional prediction module. After processing through several convolutional layers, the x and y coordinates of the four points are output as follows: The formula is as follows: (1) in, To output the set of x and y coordinates of the four points, This represents the point prediction module.

[0042] Considering the limited effective feature responses in low-contrast scenes, the point set obtained through direct regression is prone to bias. This invention introduces a shrunken prior box obtained by shrinking the ground truth bounding box as a geometric structure prior to provide explicit initial positional constraints: Given a GT box Introducing a priori box shrinkage factor This is used to adjust the degree of shrinkage of the prior box relative to the original ground truth box. The shrunken prior box. Defined as: (2) By leveraging the prior knowledge provided by the shrinking prior box, the corner coordinates of the shrinking prior box are fused with the initial predicted point coordinates to correct the point set distribution, ensuring that the point set still maintains a reasonable spatial distribution, and thus guiding the shrinking of the core point set. The four vertices are restricted to the straight lines containing the four boundaries of the contracted prior box, resulting in the final set of contracted guiding core points. Generation: (3) Where M is the fusion function, , , , To shrink the corner coordinates of the prior box.

[0043] Shrinking the core point set P The four vertices are connected in sequence to form the core prior quadrilateral, which can effectively represent the core structure of the low-contrast target and provide a stable foundation for subsequent optimization.

[0044] (3) Core Prior Quadrilateral Optimization Based on Gaussian Distribution Consistency In low-contrast scenes, the distribution of point sets is easily affected by background noise and can diverge. To address this, this invention models both the core prior quadrilateral and the contraction prior box as two-dimensional Gaussian distributions. By constraining the consistency of their distributions, the contraction guides the core point set to converge towards the target core region defined by the contraction prior box. This leverages the compact geometry of the contraction prior box to enhance the distribution stability of the contraction-guided core point set in low-contrast conditions. Experimental results show that the generated core prior quadrilateral can more accurately focus on high-contrast regions within the target, such as bright areas in dark scenes, thus extracting more effective discriminative information while reducing interference from low-contrast background regions. A schematic diagram of this design is shown below. Figure 2 As shown.

[0045] First, the core prior quadrilateral is fitted, and its Gaussian distribution is: The mean is directed towards quantity Covariance Matrix The calculation formula is as follows: (4) (5) and These are the x and y coordinates of each contraction guide core point. and They are The horizontal and vertical coordinates.

[0046] Similarly, the present invention utilizes The coordinates will also transform the shrinking prior box into a Gaussian distribution. The mean was calculated. Covariance : Since the four vertices of the core prior quadrilateral are constrained to the four sides of the contracted prior box, and each side intersects with the ground truth (GT) box, two intersection points are generated on each side, resulting in a total of eight intersection points. Among these eight intersection points, there exist exactly four vertex combinations that satisfy both the consistency with the coordinate distribution of the contracted prior box and the fact that its minimum bounding rectangle is exactly equal to the GT box—this is precisely the correct prediction result expected by this invention. The Gaussian distributions corresponding to these four combinations are consistent with the Gaussian distribution of the contracted prior box. Therefore, by approximating the contracted prior box with the Gaussian distribution of the constrained point set, the matching degree between the predicted box and the true GT box can be effectively improved.

[0047] To measure the similarity between the core prior quadrilateral distribution and the expected distribution of the contracted prior box, the following method is used: Hellinger distance is used as a measure of dissimilarity. The range of Hellinger distance is [0,1], with smaller values ​​indicating greater dissimilarity between the two distributions. The more similar they are. For two Gaussian distributions... and The Hellinger distance can be calculated using the Bhattacharyya coefficient. First, calculate their Bhattacharyya coefficients. : (6) Then, Hellinger distance It can be calculated using the Bhattacharyya coefficient: (7) Finally, the Gaussian distribution similarity loss is defined as: (8) By minimizing To optimize the shrinkage guide core point set prediction module This allows the core prior quadrilateral determined by the generated point set to fit more closely in spatial distribution to the target core region defined by the contracted prior box, especially the region with relatively high contrast. This enables more complete extraction of existing target information in the high-contrast region, while also allowing the point set to better fit the shape and contour of the target.

[0048] (4) Target regression based on the minimum bounding rectangle of the shrinking core point set Although cohesion optimization enhances the ability of the shrinking guided core point set to focus on the core region of the target, it is still necessary to ensure that the detection box covers the entire target in low-contrast scenes. Therefore, this invention introduces global coverage optimization based on minimum bounding rectangle regression. For ease of representation, The contraction-guided core point in the middle is represented as .

[0049] This invention employs the min-maximum method to select the bounding rectangle of the contracted guiding core point set as the prediction box. This method finds the minimum and maximum values ​​of the x-coordinates of the four points in the contracted guiding core point set as the left and right boundaries of the bounding box, and the minimum and maximum values ​​of the y-coordinates as the top and bottom boundaries, thus obtaining the prediction box B. (9) in The transformation process is fully differentiable and supports end-to-end gradient backpropagation. The difference between the predicted and ground truth boxes is supervised by a regression loss, thereby achieving joint optimization from shrinking the core point set generation to the final detection box output.

[0050] 3. Loss Function Design To achieve synergistic optimization of core point set generation, distribution constraints, and detection accuracy guided by shrinkage, this invention constructs a multi-objective joint loss function system. The total loss function is composed of a weighted average of classification loss and overall regression loss. (10) The definitions of each loss are as follows: (1) Classification loss

[0051] Focal Loss is employed to mitigate the extreme imbalance between positive and negative samples in low-contrast scenes. Its form is as follows: (11) This represents the model's predicted probability for the correct category. As a class balance factor, It is a regulating factor.

[0052] (2) Overall regression loss

[0053] The overall regression task is supervised by two parts of loss: Gaussian distribution similarity loss. and bounding box regression loss .

[0054] Bounding box regression loss The design approach involves comparing the transformed predicted bounding box B with the true target bounding box G, and using a robust SmoothL1 loss function for supervision. The specific definition of SmoothL1 loss is as follows: (12) in, The difference between the predicted and actual coordinate values. This is the threshold parameter.

[0055] The Gaussian loss has been explained in detail in step three of the network setup section. The total regression loss is the weighted sum of the two losses: (13) in, and To balance the weights.

[0056] 4. Model Training and Validation This invention designs a complete model training and verification process to ensure that the proposed shrinkage-guided core point set representation target detection method has the best performance and the strongest generalization ability in low-contrast scenes.

[0057] (1) Training process During training, the model was fed low-light images along with their corresponding ground truth bounding boxes and class labels in batches. Forward propagation was used to calculate predictions and various losses, while backpropagation was used to update network parameters. A stochastic gradient descent optimizer was used with a momentum of 0.9 and a weight decay of 0.0001. The initial learning rate was set to 0.001, and a cosine annealing scheduling strategy was employed, decreasing by a factor of 0.1 at epochs 18 and 22. The total training epochs were 100, with a batch size of 8. All experiments were implemented using the PyTorch deep learning framework and conducted in a distributed manner on an NVIDIA GeForce RTX 3090 GPU.

[0058] To enhance the model's adaptability to low-contrast scenes, a multi-scale training strategy was employed during training, with the shorter side of the input image randomly sampled from the interval [480, 960]. Simultaneously, data augmentation techniques such as random horizontal flipping and color dithering were used to improve the model's generalization ability without compromising the original low-contrast degradation characteristics. During training, the trends of classification loss, regression loss, and total loss curves were recorded to monitor the model's convergence status and gradient stability in real time.

[0059] (2) Verification process During the validation phase, the model performs a single forward inference based solely on the input image, outputting the bounding box coordinates and class confidence score. To comprehensively evaluate the model's detection performance in low-contrast scenes, this invention employs a common evaluation metric in the object detection field: mean accuracy (mAP@0.5).

[0060] The mean accuracy is the arithmetic mean of the average accuracy of each category, calculated using an IoU threshold of 0.5. It comprehensively reflects the model's overall performance in localization and classification tasks. This metric exhibits good robustness to boundary ambiguity, scale variations, and class imbalance, making it a core evaluation benchmark for low-contrast detection tasks.

[0061] (3) Experimental results Table 1. Detection results of different models on the ExDark dataset

[0062] As shown in Table 1, the quantitative experimental results on the ExDark low-light dataset demonstrate that the target detection method based on the contracted guided core point set representation proposed in this invention achieves a mAP@0.5 of 71.1%, significantly outperforming mainstream comparison methods. Compared with methods such as RetinaNet (65.5%), GFLV (68.6%), MAET (63.7%), and MADet (69.1%), this invention improves the overall detection accuracy by 5.6%, 2.5%, 7.4%, and 2.0%, respectively. Especially in categories that are prone to missed detection and difficult to locate under low contrast conditions, such as Bottle, Table, and People, this invention achieves AP values ​​of 70.4%, 43.4%, and 64.8%, respectively, representing improvements of over 4–16 percentage points compared to the optimal baseline. This verifies the significant advantages of the contracted guided core point set representation and Gaussian optimization mechanism under extreme degradation conditions.

[0063] like Figure 3 As can be seen from the visualization, the single-stage target detection algorithm designed in this invention can still maintain a high accuracy rate even under very low illumination and with very little effective information, thus avoiding the occurrence of missed detection events.

[0064] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A low-contrast target detection method based on shrinking guided core point set representation, characterized in that, Specifically, the following steps are included: S1. Dataset Construction: Based on publicly available low-light target detection benchmark datasets, the datasets are preprocessed to obtain low-light images to be detected. S2. Network Structure Construction: A single-stage target detection network model for low-contrast scenes is proposed, which follows the classic Backbone-FPN-Head paradigm. The detection head of the network model adopts a target regression method based on shrinking guided core point set representation to replace the traditional anchor box regression or key point estimation mechanism. S3. Loss Function Design: Construct a multi-objective joint loss function system to jointly constrain the generated results at the geometric structure and semantic alignment levels, so as to achieve synergistic optimization of point set generation, distribution constraints and detection accuracy; S4. Model Training and Validation: Based on S1~S3, design a complete model training and validation process; During the training phase, the model is input with low-light images and their corresponding ground truth bounding boxes and category labels in batches. The prediction results and various losses are calculated through forward propagation, and the network parameters are updated using the backpropagation algorithm. Multi-scale training strategies and data augmentation techniques are employed to enhance the model's adaptability and generalization. During the validation phase, the model relies on the input image to perform a single forward inference and outputs the coordinates of the detection box and the category confidence score.

2. The low-contrast target detection method based on shrinking guided core point set representation according to claim 1, characterized in that, The dataset preprocessing described in S1 specifically includes: All images in the dataset are uniformly resized to a fixed input size and then normalized using mean-variance. A random partitioning strategy is adopted, with 80% of the images used for training and 20% for testing, to ensure a balanced distribution of samples for each category under different lighting conditions, providing a stable and reliable data foundation for model training and performance evaluation.

3. The low-contrast target detection method based on shrinking guided core point set representation according to claim 1, characterized in that, The single-stage object detection network model for low-contrast scenes described in S2 specifically includes: Backbone Network and Feature Pyramid Module: A pre-trained ResNet-50 is used as the backbone network to extract multi-scale deep features from the input image, and a multi-scale feature map is generated through the feature pyramid network to provide rich semantic and spatial information for subsequent point set regression. The shrinking guided core point set generation module based on shrinking prior boxes is used to receive the feature map output by the feature pyramid network and output the initial shrinking guided core point coordinates through a lightweight convolutional prediction head. The prior box obtained by shrinking the real labeled box is introduced as a strong geometric constraint. By fusing the initial predicted point set coordinates with the shrinking prior box, the four points of the final shrinking guided core point set are strictly constrained to the straight lines of the four boundaries of the shrinking prior box. Connecting the four points forms a core prior quadrilateral to enhance the feature focusing ability of the target core region. The core prior quadrilateral optimization module based on Gaussian distribution consistency models the core prior quadrilateral and the shrinking prior box as two-dimensional Gaussian distributions respectively, and achieves fine-tuning of the point set by minimizing the distribution difference between the two. The target regression module based on the minimum bounding rectangle of the shrinking guide core point set: The minimum bounding rectangle method is used to construct the prediction box by taking the minimum and maximum horizontal and vertical coordinates of the shrinking guide core point set, so as to achieve joint optimization from the generation of the shrinking guide core point set to the output of the final detection box.

4. The low-contrast target detection method based on shrinking guided core point set representation according to claim 1, characterized in that, The multi-objective joint loss function system includes classification loss and overall regression loss, and achieves comprehensive constraints through a weighted summation. The classification loss adopts Focal Loss to alleviate the problem of extreme imbalance between positive and negative samples in low-contrast scenes, so that the model pays more attention to the difficult-to-classify samples and improves the classification confidence and discrimination ability. The overall regression loss is composed of a weighted sum of the bounding box regression loss and the Gaussian distribution similarity loss, with the weights determined through cross-validation tuning.

5. The low-contrast target detection method based on shrinking guided core point set representation according to claim 4, characterized in that, The bounding box regression loss uses Smooth L1 loss to constrain the coordinate difference between the predicted box and the ground truth box, so as to ensure the localization accuracy of the detected box and reduce the impact of outliers on gradient updates. The Gaussian distribution similarity loss, based on Hellinger distance, is used to constrain the contraction of the core prior quadrilateral and guide the core point set distribution to converge toward the contracted prior box, thereby enhancing the spatial stability and geometric compactness of the point set under low contrast.

6. The low-contrast target detection method based on shrinking guided core point set representation according to claim 1, characterized in that, During the training phase, a stochastic gradient descent optimizer was used with a momentum of 0.9 and a weight decay of 0.0001. The initial learning rate was set to 0.001, and a cosine annealing scheduling strategy was adopted, with the learning rate decreasing by a factor of 0.1 in the 18th and 22nd rounds. The total number of training rounds is 100, and the batch size is 8.

7. The low-contrast target detection method based on shrinking guided core point set representation according to claim 1, characterized in that, During the verification phase, the average accuracy was used as the performance evaluation metric for the model's detection in low-contrast scenarios.

8. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the instruction, program, code set, or instruction set is loaded and executed by the processor to implement a low-contrast target detection method based on a shrinking guided core point set representation as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, which is loaded and executed by a processor to implement a low-contrast target detection method based on a contracted guide core point set representation as described in any one of claims 1-7.

Images