Target detection method in multi-scale self-adaptive low-light environment based on intensity and region specificity

Abstract

The invention belongs to the field of deep learning network target detection, and discloses a multi-scale adaptive low-light environment target detection method based on intensity and region specificity. According to the method, the Laplace multi-scale feature enhancement network structure is designed, so that the target detection performance in a low-light environment is remarkably improved. According to the LaMFE, through multi-scale feature fusion, information of different scales is fully utilized, a low-light image is finely optimized, and meanwhile, an efficient residual connection mechanism is adopted, so that the feature extraction efficiency and precision are improved. The design not only improves the visualization quality of the low-light image, but also significantly improves the robustness and accuracy of the target detection model in a complex low-light environment, so that the target detection model has excellent performance in wide application scenes.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning network target detection, and in particular relates to a low-light intensity target detection method based on multi-scale flexible adaptation. Background Art

[0002] Object detection is a critical task in computer vision and plays a key role in a wide range of applications such as surveillance, autonomous driving, and military operations. The ability to accurately identify and localize objects in images is critical to the functionality and safety of these systems. However, a significant challenge that hinders the performance of traditional object detection algorithms is the presence of low-light environments.

[0003] In low-light environments, images are often affected by noise superposition, reduced contrast, and loss of color information, making it difficult for traditional detection methods to perform effectively. These low-light environments are common in surveillance, such as nighttime security cameras that often operate in limited ambient light, or in autonomous driving scenarios at dusk or even at night, where the surrounding scene is low in brightness. In addition, military operations often occur in low-visibility conditions, further emphasizing the need to study robust low-light target detection techniques.

[0004] Low-light image enhancement methods can be roughly divided into traditional methods and deep learning methods. Traditional methods mainly include methods based on distribution mapping and model optimization. Although traditional methods have improved image brightness and quality to a certain extent, methods based on distribution mapping often ignore the lighting requirements of low-light images and the intrinsic connection between pixels, which may lead to color distortion and abnormal details. Model optimization-based methods have limitations in prior design and are prone to color shift, low exposure, ghosting and noise problems.

[0005] The low-light image enhancement method based on deep learning started in 2017 and has become the mainstream method for low-light image enhancement. In general, from the perspective of the purpose of implementation, the low-light image enhancement method based on deep learning can be roughly divided into two categories: methods for brightness enhancement and methods for combined brightness enhancement and noise removal. The first type of method increases the brightness of the image to show more detailed information of the image. The second type is the existing mainstream work, which combines brightness enhancement and noise removal, and solves the brightness and noise problems in low-light images at the same time through a deep learning network, which can comprehensively improve the brightness and visual quality of the image.

[0006] Current low-light image enhancement technology has gone through a development from traditional model design to data-driven deep learning. In terms of learning mechanisms, it is gradually moving towards semi-supervised / unsupervised learning and is no longer limited to fully supervised learning. At the same time, the evaluation method is also changing, from focusing only on visual quality to focusing more on the performance of downstream high-level visual tasks.

[0007] Low-light object detection depends on the further development of low-light enhancement technology, and there are still many problems in the existing research on low-light object detection. There are two main methods for object detection under low-light conditions: one is to first enhance the low-light image and then perform object detection on the enhanced image; the other is to directly optimize the enhancement network with object detection as the goal without generating an enhanced image.

[0008] The first method, often referred to as enhancement before detection, first enhances the image under low-light conditions to improve image quality and visibility, and then feeds the enhanced image into the object detector for object detection. However, this method may introduce additional noise or distortion during the enhancement process, affecting the detection performance.

[0009] The second method is to optimize the enhancement network specifically to achieve better object detection performance. Unlike traditional image enhancement algorithms, the optimization goal of this method is to minimize the detection loss, and object detection can be performed even without generating enhanced images. The advantage of this method is that it can directly optimize the object detector while avoiding the additional noise or distortion that may be introduced by the enhancement process.

[0010] Recently, there have been a series of advances in low-light object detection with enhanced networks directly targeting object detection. Liu et al. proposed IA-YOLO, which includes a differentiable image processing (DIP) module that takes into account the adverse weather conditions of the YOLO detector, and its parameters are predicted by a small convolutional neural network (CNN-PP), which jointly learns CNN-PP and YOLOv3 to ensure that CNN-PP can learn appropriate DIP to enhance images for detection in a weakly supervised manner. Cui et al. proposed MAET, which learns intrinsic visual structure in a self-supervised manner by encoding and decoding realistic lighting degradation transformations, while considering physical noise models and image signal processing (ISP). Based on this representation, bounding box coordinates and categories are decoded to achieve the object detection task. In order to avoid excessive entanglement of the two tasks, the orthogonal tangent rule is imposed to decouple the object and degraded features of MAET. This forms a parameterized manifold on which multi-task predictions can be geometrically formulated by maximizing the orthogonality between the tangents output by each task.

[0011] Multi-scale feature extraction and fusion is a key image processing technology that aims to extract features from different scales and effectively fuse them to improve the performance of image analysis tasks. Through multi-scale feature extraction, the global information and local details of the target can be captured, and it can adapt to targets of different sizes, shapes and positions. In addition, multi-scale feature fusion can make full use of the complementarity of features at different scales, combine the detailed expression of low-level features with the semantic information of high-level features, and thus generate more robust feature representations. This technology is widely used in tasks such as target detection and image segmentation, and significantly improves the adaptability of the algorithm to complex scenes. In low-light target detection, the role of multi-scale feature extraction and fusion is particularly important. In low-light environments, the degradation of image quality poses a huge challenge to target positioning and recognition. Related studies have shown that multi-scale technology can effectively improve low-light detection performance. For example, the DENet proposed by Qin et al. uses the Laplacian pyramid to decompose the input image into different frequency bands, respectively enhances the global brightness of the low-frequency component and the detailed information of the high-frequency component, and improves the expressiveness of the low-light image. The FeatEnHancer module proposed by Hashmi et al. combines the multi-head attention mechanism to guide the fusion of multi-scale features through task-related losses, thereby improving the quality of feature expression. These methods verify the importance of multi-scale features in capturing complex scene information and improving low-light detection accuracy.

[0012] Inspired by these studies, we proposed a new solution to improve the performance of low-light object detection. We designed a plug-and-play module - the Laplacian Multi-Scale Feature Enhancement (LaMFE) network structure. LaMFE can locally and adaptively enhance the brightness of the background area with the help of the overall image intensity, while limiting the noise interference introduced by the enhancement to the target area, thereby significantly improving the object detection performance. In addition, common enhancement methods have different effects on different degrees of low-light environments: when the light is not particularly dark, the enhancement method may be ineffective or even have side effects; while in extremely low-light conditions, some enhancement algorithms are effective. To address this problem, our method is based on the multi-scale adaptive low-light environment object detection method with intensity and region specificity, which can flexibly select the best solution and achieve flexible low-light object detection. Summary of the invention

[0013] In response to the above problems, we propose a Laplacian Multi-Scale Feature Enhancement (LaMFE) network structure, which significantly improves the performance of target detection in low-light environments through multi-scale feature fusion and efficient residual connection. As a plug-and-play module, LaMFE can be directly applied to low-light images, locally enhance the background part, and support multiple cycles until the desired effect is achieved. This flexibility enables LaMFE to significantly improve the accuracy and reliability of low-light target detection in various practical applications.

[0014] The technical solution of the present invention:

[0015] A multi-scale adaptive low-light environment target detection method based on intensity and region specificity, comprising the following steps:

[0016] Step 1: Obtain low-light images. The images can come from image acquisition devices in low-light environments such as surveillance cameras and night vision devices. The low-light images are used as the initial input for target detection and background enhancement. After image enhancement, feature extraction, and target positioning, the target detection performance under low-light conditions is improved.

[0017] Step 2: Input the processed low-light image into the pre-selected target detection model, and use the feature extraction module in the target detection model to extract multi-scale features; by analyzing the multi-scale features, locate the potential target object in the processed low-light image, and output the detection result, including the target's bounding box, category label and confidence score (0-1). The confidence score is used to indicate the reliability of the target detection. The preliminary target detection provides a rough position reference of the target area for the subsequent background enhancement step.

[0018] The input low-light image I is roughly detected, and the preliminary information of the low-light image I is extracted using the target detection model M. The specific process is expressed as D = M (I), where D is the detection result, including the target detection box B, the target category label C and the confidence score S; the output form of each detection result is D i = {B i ,C i ,S i}, where B i ={x i ,y i ,w i ,h i} represents the location information of the target detection box, C i represents the target category label, S i It is the confidence score of the target detection model for the detection result, ranging from 0 to 1;

[0019] Through coarse detection, we can quickly locate the target area in the low-light image and output the rough position and category information of the target. The detection box B provides the spatial distribution of the target, the target category label category C clarifies the nature of the target, and the confidence score S reflects the reliability of the detection. This information provides an important basis for subsequent background enhancement, ensuring that the target area can be retained without interference during the enhancement process while optimizing the brightness characteristics of the background. Coarse detection is not only a key step in the initial screening of targets, but also provides a decision-making basis for multiple rounds of optimization of the system, helping to adjust the background enhancement strength to gradually improve the target detection performance.

[0020] Step 3: According to the rough detection results, the target area in the low-light image I is determined, and the surrounding part is defined as the background area; the target area and the background area are accurately divided by coordinate calculation and image segmentation technology, and only the background area is enhanced; a Laplace multi-scale feature enhancement network structure LaMFE is constructed, including a Laplace decomposition structure, a Laplace reconstruction structure, and a ResNetBlock module for feature extraction; wherein the ResNetBlock module sequentially executes two Block modules to extract features. Assuming that the input feature is x, the Block module sequentially performs group normalization, Swish activation function, Dropout, and a 3×3 convolution layer on the feature x to obtain the feature x′, and then adds the feature x′ to the residual of the input feature x; in the Laplace decomposition structure, the decomposition level m is set to 3, and the low-light image I is decomposed by the Laplace pyramid to convert it into three feature maps of different scales I1, I2, and I3. The decomposition formula is as follows:

[0021] I1=IU(D(G(I))), I2=I1-U(D(G(I1))), I3=D(G(I2))

[0022] Among them, G represents the Gaussian blur process, D represents downsampling, and U represents upsampling; the three feature maps of different scales after decomposition cover feature information of different scales: feature map I1 is a reconstructed version after subtracting blur from the low-light image I, retaining the high-frequency detail information in the low-light image I, including texture and edge; feature map I2 is obtained by performing the same processing on feature map I1 to extract lower-scale detail features; feature map I3 is obtained by directly performing Gaussian blur and downsampling on feature map I2 to reflect the global low-frequency features. After obtaining the feature map, the feature fusion process is as follows:

[0023] (1) Downsample the feature map I1 to the same scale as the feature map I2, recorded as feature map I′1, and input it into the ResNetBlock module together with the feature map I2 to complete feature fusion and enhancement, and obtain feature map I′2;

[0024] (2) Downsample the feature map I′2 to the same scale as the feature map I3, record it as feature map I″2, and input it into the ResNetBlock module together with the feature map I3 to obtain the fused enhanced feature map I′3;

[0025] Finally, in the Laplacian reconstruction structure, the I′3, I′2, and I1 feature maps are upsampled layer by layer to restore the image to the original resolution, generating the preliminary output enhanced_features; then, the enhanced_features are residually added to the low-light image I to retain the global brightness information of the original image, and the final output I′ is obtained; this design not only enhances the detail performance but also improves the global consistency. The Laplacian multi-scale feature enhancement network structure LaMFE is used to enhance the background of the low-light image, improve its brightness and contrast, while keeping the target area unchanged to avoid feature interference;

[0026] Based on the output result of rough detection D i = {B i ,C i ,S i}, accurately divide the target area and background area in the low-light image I; specifically, through the target detection frame B i The coordinates (x i ,y i ,w i ,h i ), the low-light image I is divided into a target area and a background area; the target area consists of the part covered by the detection frame, and the background area is the image part outside the target area; it is expressed as:

[0027] R 背景 =R(I)\R 目标

[0028] Among them, R 目标 Indicates the target area, R(B i ) is the target detection box B i The corresponding area, R(I) is the pixel set of the low-light image I, R 背景 is the remaining background area;

[0029] The division process is completed through coordinate calculation and image segmentation technology; first, the pixel range of each target area is accurately located according to the boundary coordinates of the target detection frame; then, these target areas are subtracted from the low-light image I through logical operations to obtain the background area; this process provides a clear area range for subsequent background enhancement operations, allowing the enhancement process to focus on the background part without affecting the feature integrity of the target area. Through the division of the target area and the background area, the brightness of the background area is adjusted in a targeted manner while keeping the features of the target area unchanged; this precise separation not only improves the effectiveness of background enhancement, but also lays the foundation for the improvement of the final target detection performance.

[0030] Step 4: Send the enhanced low-light image to the target detection model in step 2 again for processing, and observe the target detection effect after the background area is enhanced; through the detection confidence feedback, evaluate whether the current enhancement has an optimization effect on target detection;

[0031] Assuming that the image after background enhancement is I′, the process of the coarse target detection algorithm is expressed as:

[0032] D′=M(I′)

[0033] Among them, D′ represents the new detection result, including the updated target detection box B′ i , target category label C′ i and confidence score S′ i ;

[0034] By comparing D′ and D, the effectiveness of background area enhancement is evaluated; if the target detection box B′ i More accurately covers the target area, while the confidence score S′ i If the confidence score S′ is improved, it indicates that the current background area enhancement strategy is effective; otherwise, i If the background image drops or the detection frame shifts, you may need to adjust the intensity or method of background enhancement. The enhancement effect is evaluated by the following indicators:

[0035] ΔS i =S′ i -S i

[0036] If ΔS i >0, background area enhancement is effective; otherwise, adjustment is required;

[0037] The purpose of repeated coarse detection is to gradually optimize the visibility and detection performance of the target in multiple rounds of enhancement and detection. The feedback from each round of coarse detection will affect the strength or range of the next round of background enhancement. For example, in some low-light scenes, it may be necessary to strengthen the background brightness enhancement factor, while in other scenes, it may be necessary to more finely control the boundaries of the enhancement area.

[0038] This step ensures that the iterative optimization of the system can dynamically adapt to the needs of different scenarios, continuously improve the background enhancement and target detection performance, and lay a solid foundation for the final accurate detection.

[0039] Step 5: Repeat the loop process of background area enhancement and coarse target detection for n iterations in total; in each iteration, dynamically adjust the intensity of background area enhancement according to the results of the previous round of detection; for example, if the detection results show that excessive enhancement causes interference with target features, then appropriately reduce the enhancement intensity. Through multiple rounds of optimization, gradually improve image quality and detection performance until the detection requirements are met;

[0040] Specifically: By comparing the k-th round rough detection result D (k) , according to the target detection box Target category label and confidence score changes, judging whether it is necessary to optimize the background area enhancement parameters;

[0041] Specific strategies for adjusting background area enhancement include:

[0042] 1) Increase or decrease the strength of the reinforcement:

[0043] For the kth round of detection, compare the target confidence score of the previous round and the current confidence score change If the confidence score Increase that is ΔS i >0, indicating that the accuracy of target detection has improved, and the current background area enhancement strength may be effective, so the current enhancement strategy is maintained or strengthened;

[0044] If the confidence score increases: maintain the current background area enhancement strength, that is, maintain the brightness increase factor α (k) constant;

[0045] If the confidence score decreases, it means that the current background area is too enhanced, which may have excessively affected the characteristics of the target, thereby reducing the detection effect. In this case, reduce the brightness enhancement factor α (k) ;The adjustment formula of the brightness enhancement factor is:

[0046]

[0047] Wherein, Δα is the increment of the brightness factor, which is a preset constant or a dynamically calculated value, and is adjusted according to the magnitude of the confidence change;

[0048] 2) Optimize and enhance the area:

[0049] If the target detection box If there is a deviation, the target area R is re-divided accurately. 目标 and background area R 背景 , avoiding interference with the target area, while strengthening the optimization of the background area of ​​low-light images;

[0050] During multiple rounds of iterations, the background enhanced image I′(k+1) and the detection result D in each round (k+1) All depend on the output results of the previous round:

[0051]

[0052] This iterative approach enables the system to be gradually optimized based on detection feedback, ultimately achieving efficient and stable target detection in low-light environments. The core of the fifth step is to achieve the best balance between background brightness adjustment and target detection performance through multiple rounds of enhancement and detection loops. By analyzing the detection results of each round and flexibly adjusting the enhancement strategy, the system can effectively cope with different low-light scenes, ensuring clear target features and accurate detection results.

[0053] Step 6: After completing n iterations, input the optimized image into the precise target detection model for final detection; perform processing on the enhanced image to obtain more precise and stable target detection results; use the high-performance model to process on the enhanced image to obtain accurate and stable target detection results suitable for actual application scenarios.

[0054] The final enhanced image I′ (n) Input a high-precision target detection algorithm M final Perform final testing to obtain accurate test results.

[0055] After completing n enhancement iterations, the background brightness and features have been fully optimized, the target area is clearly visible, and the contrast between the background and the target is significantly improved. At this point, the final enhanced image I′ (n) Input a high-precision target detection algorithm M final Perform final detection to obtain accurate detection results. Compared with coarse detection, precise detection algorithm can better capture target details and improve positioning accuracy and classification confidence.

[0056] Furthermore, the high-precision target detection algorithm M final Choose Faster R-CNN or RetinaNet. These models have the ability of multi-stage optimization. First, for the input image I′ (n) Perform feature extraction and target analysis, and process multi-scale targets through convolution feature pyramid; secondly, further adjust the position and size of the detection frame, and recalculate the target confidence of the target category; the detection process is described as: I′ (n) Input high-precision target detection algorithm M final , output the final target detection box D final , including target detection box, target category label and confidence score.

[0057] Final test result D finalIt includes all detection targets in low-light environments, with small positioning errors and high category recognition accuracy. To ensure the detection effect, the comprehensive performance of the enhancement module LaMFE (Laplacian multi-scale feature enhancement network structure) and the target detection algorithm (YOLOv3 and TOOD) is verified by comparing with standard detection results or manual annotations. This stage fully utilizes the enhanced image characteristics and the capabilities of the high-precision detection algorithm, significantly improving the target detection performance in low-light environments and meeting the needs for efficient and reliable detection in practical applications.

[0058] Furthermore, the final test result D final Contains all detected targets in the background area of ​​the low-light image; to ensure the detection effect, the comprehensive performance of the enhanced model LaMFE and the precise target detection model is verified by comparing with the standard detection results or manual annotations.

[0059] Furthermore, the average brightness value of the low-light image is less than 30, and the pixel value range is 0-255.

[0060] Furthermore, in step 2, the target detection model is YOLOv3.

[0061] Furthermore, in step 6, the precise target detection model is YOLOv8 or TOOD.

[0062] Beneficial effects of the present invention: By designing the Laplace Multi-Scale Feature Enhancement (LaMFE) network structure, the present invention significantly improves the target detection performance in low-light environments. LaMFE uses multi-scale feature fusion to fully utilize information at different scales to carefully optimize low-light images, while adopting an efficient residual connection mechanism to improve the efficiency and accuracy of feature extraction. This design not only improves the visualization quality of low-light images, but also significantly improves the robustness and accuracy of target detection models in complex low-light environments, making it perform well in a wide range of application scenarios.

[0063] The core advantage of the LaMFE module is that it is combined with coarse detection to form a multi-round cyclic enhancement mechanism, focusing on local optimization of the background area while keeping the detection target unchanged. During the cycle, the coarse detection module provides clear guidance for background enhancement by preliminarily identifying the target area and providing location and category information, so that each round of enhancement can focus on improving background brightness and contrast. At the same time, coarse detection can also evaluate the enhancement effect, dynamically adjust the intensity and range of background enhancement according to the detection results, and ensure that the optimization process gradually converges. This effectively avoids the defect of traditional methods that are prone to over-processing, not only ensuring the integrity of the detection target, but also significantly reducing the side effects that may be caused by the enhancement operation.

[0064] In addition, LaMFE adopts a plug-and-play modular design and can be flexibly integrated into the existing target detection framework. With the synergy of coarse detection and multiple rounds of enhancement, LaMFE achieves high adaptability and can dynamically adjust the enhancement strategy and number of cycles according to the specific scene to ensure that the final enhancement effect meets the expectations. This flexibility and reliability enable LaMFE to show significant application value in image processing and target detection tasks in low-light environments, providing an efficient and stable solution for related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A detailed training framework structure diagram of the target detection method in a low-light environment based on intensity and area adaptation provided by the present invention;

[0066] Figure 2 This is a flow chart of the target detection method in a low-light environment based on intensity and area adaptation provided by the present invention. DETAILED DESCRIPTION

[0067] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0068] like Figure 1 As shown, the present invention provides a multi-scale adaptive target detection method in low-light environment based on intensity and region specificity.

[0069] like Figure 2 As shown, the specific implementation is as follows:

[0070] 1. Obtain low-light image training data

[0071] First, obtain low-light image training data pairs. Images can be obtained by taking images under low-light conditions and annotating them yourself, or by using public low-light datasets such as Exdark or DarkFace. Next, the image size is cropped to 408 pixels on the short side and 1333 pixels on the long side to ensure that the input image size is consistent. Then, a series of data augmentation operations are performed, including random cropping (randomly selecting an area from the original image for cropping), random rotation (randomly rotating the image by a certain angle, such as between -10 degrees and +10 degrees), random flipping (flipping the image horizontally or vertically), and random adjustment of brightness and contrast. Through these enhancement operations, the robustness of the model can be improved, overfitting can be avoided, and its target detection performance in low-light environments can be enhanced.

[0072] 2. Preliminary target detection

[0073] In object detection in low-light image processing, we first input the low-light image I into a pre-trained object detection model M (such as YOLOv3). First, we normalize it, map the pixel values ​​to the [0,1] range, and input the processed image into the model M for object detection.

[0074] The model M outputs a preliminary detection result D. For example, suppose the model detects two people, D1 and D2, where:

[0075] D1 = {B1, C1, S1}, where B1 = {x1 = 50, y1 = 60, w1 = 100, h1 = 200}, indicating the position of the detection box of the first target, C1 = "person", indicating that the target category is "person", S1 = 0.85, indicating that the model's confidence in the target is 85%;

[0076] D2 = {B2, C2, S2}, where B2 = {x2 = 150, y2 = 160, w2 = 120, h2 = 210}, indicating the position of the detection box of the first target, C2 = "person", indicating that the target category is "person", and S2 = 0.90, indicating that the model's confidence in the target is 90%;

[0077] Next, we filter the detection results and set the confidence threshold T = 0.8. Since S1 = 0.85 and S2 = 0.90 are both above the threshold, we retain these two detection results. Then we apply non-maximum suppression (NMS), assuming that the IoU between the two boxes is 0.4 (less than the set 0.5 threshold), so both are retained.

[0078] The final test results are:

[0079] D′={D1={B1,C1,S1},D2={B2,C2,S2}}

[0080] This means that the model successfully detected two targets, two objects of the "person" category, and output the detection box, category label and confidence score of each target.

[0081] 3. Background area determination and enhancement

[0082] Next, the target detection box output by the preliminary detection model YOLOv3 is B1 = {x1 = 50, y1 = 60, w1 = 100, h1 = 200}. This box represents a rectangular area with a width of 100 and a height of 200 starting from the image coordinates. This area contains a person and is the target area R 目标 The part outside the target area is defined as the background area.

[0083] We precisely locate the target region by the coordinates of the target box, subtract it from image I, and the remaining part is the background region. The target region contains pixels from (50,60) to (150,260), while the background region is all pixels outside the target box.

[0084] Next, we use the LaMFE model to quantize the background region R 背景 Perform enhancement processing. Assume that after LaMFE enhancement, the brightness of the background is increased by 30% and the contrast is increased by 25%, making the background part clearer and more visible. 目标 Keep it as is, avoiding any feature interference.

[0085] In the final processed image, the background area becomes bright and clear, while the target area retains its original features. This method improves the quality of low-light images while ensuring the accuracy of target detection.

[0086] 4. Enhanced target detection and feedback

[0087] Next, we feed the enhanced image I′ into the coarse object detection model for a second detection to observe whether background enhancement has an optimization effect on object detection.

[0088] First, the enhanced image I′ is input into the target detection model M to obtain the new detection result D′, which is expressed as:

[0089] D′=M(I′)

[0090] Assume that after detection, the new target detection box is B′1={x′1=48,y′1=58,w′1=102,h′1=202}, the target category label is C′1, and the confidence score is S′1=0.92. Compared with the previous detection result B′1={x′1=48,y′1=58,w′1=102,h′1=202}, we can see that the position and size of the target box have been adjusted, and the confidence score has also been improved.

[0091] To evaluate whether the background enhancement is effective, we calculate the difference in confidence scores ΔS1:

[0092] ΔS1=S′1-S1=0.92-0.85=0.07

[0093] Due to ΔS1, it shows that background enhancement improves the model's confidence in the target, verifying the effectiveness of background area enhancement. In addition, the target box B′1 also covers the target area more accurately, which further shows that background enhancement makes the features of the target area more prominent and the detection results are optimized.

[0094] 5. Multiple rounds of iterative optimization

[0095] Repeat the loop process of background area enhancement and rough target detection for a total of 3 iterations. In the first round of iteration, the image after background enhancement is I′1, and the detection result output by the target detection model is Compared with the first round of detection results, the confidence score Than the first time It has increased by 0.07. Therefore, This shows that the background enhancement strategy is effective, so we decide to continue to maintain the current enhancement strategy in the second round of iterations.

[0096] In the second round of iteration, the enhanced image I′1 is input into the target detection model again for detection, and a new detection result is obtained. At this point, the position of the target box remains stable and the confidence score Compared with the previous round Increased by 0.01, calculated because It is considered that the current background enhancement strength is still valid, so the brightness enhancement factor α is maintained (2) constant.

[0097] In the third iteration, the enhanced image I′2 is used as input for detection again, and the target detection result is At this time, the position information of the target frame changes little, and the confidence score Compared with the previous round Improved by 0.01, we get At this time, due to We continue to maintain the current background enhancement strategy and increase the brightness by a factor of α (3) After multiple rounds of iterations and 3 optimizations, the position and size of the target box gradually stabilized, and the confidence score gradually increased from the initial Promote to The image quality of the target area is improved by about 10.6%. In this process, the enhancement of the background area effectively improves the brightness and contrast of the image, while keeping the characteristics of the target area unchanged. Finally, the image I′3 after n=3 iterations of optimization will be input into a higher-precision target detection model, such as YOLOv8, TOOD, etc., for final detection to obtain more accurate and stable detection results.

[0098] 6. Final target detection

[0099] Finally, after three rounds of background area enhancement and target detection optimization, the image I′ is obtained. (3) It is sent to the high-precision target detection model YOLOv8 for final detection. After processing, the target detection result is:

[0100]

[0101] The target frame The target position is accurately located, and the confidence score is 0.95, indicating that the target detection accuracy is high.

[0102] The present invention improves the accuracy of target detection in low-light images through precise background area enhancement and target detection optimization strategies. After preliminary target detection, the target area and background area are determined based on the target detection frame, and the background is enhanced using the LaMFE model to make the background brighter and clearer, while keeping the target area unchanged to avoid interference. Through multiple rounds of iterative optimization, the enhancement intensity is dynamically adjusted according to the results of each round of detection, and the detection confidence and accuracy of the target frame are continuously improved. Finally, the optimized image is input into a high-precision target detection model (such as YOLOv8) for processing to obtain more accurate and stable target detection results, which significantly improves the performance of target detection in low-light environments.

Claims

1. A multi-scale adaptive low-light target detection method based on intensity and region specificity, characterized in that: The following steps are involved: Step 1: Obtain low-light images and use them as the initial input for target detection and background enhancement. After image enhancement, feature extraction, and target localization, the target detection performance in low-light conditions is improved. Step 2: Input the processed low-light image into the pre-selected target detection model, and use the feature extraction module in the target detection model to extract multi-scale features; by analyzing the multi-scale features, locate the potential target object in the processed low-light image, and output the detection results, including the target's bounding box, category label and confidence score, to provide basic data for subsequent tasks; The input low-light image I is roughly detected, and the preliminary information of the low-light image I is extracted using the target detection model M. The specific process is expressed as D = M (I), where D is the detection result, including the target detection box B, the target category label C and the confidence score S; the output form of each detection result is D i = {B i ,C i ,S i }, where B i ={x i ,y i ,w i ,h i } represents the location information of the target detection box, C i represents the target category label, S i It is the confidence score of the target detection model for the detection result, ranging from 0 to 1; Step 3: According to the rough detection results, determine the target area in the low-light image I, and define the surrounding area as the background area; accurately divide the target area and the background area through coordinate calculation and image segmentation technology, and only enhance the background area; construct a Laplace multi-scale feature enhancement network structure LaMFE, including a Laplace decomposition structure, a Laplace reconstruction structure, and a ResNetBlock module for feature extraction; wherein the ResNetBlock module sequentially executes two Block modules to extract features. Assuming that the input feature is x, the Block module sequentially performs group normalization, Swish activation function, and Dropout on the feature x. And a 3×3 convolution layer is used to obtain feature x, and feature x is added to the residual of input feature x. In the Laplace decomposition structure, the decomposition level m is set to 3, and the low-light image I is decomposed by the Laplace pyramid to convert it into three feature maps I1, I2 and I3 of different scales. The decomposition formula is as follows: I1=IU(D(G(I))), I2=I1-U(D(G(I1))), I3=D(G(I2)) Among them, G represents the Gaussian blur process, D represents downsampling, and U represents upsampling; the three feature maps of different scales after decomposition cover feature information of different scales: feature map I1 is a reconstructed version after subtracting blur from the low-light image I, retaining the high-frequency detail information in the low-light image I, including texture and edge; feature map I2 is obtained by performing the same processing on feature map I1 to extract lower-scale detail features; feature map I3 is obtained by directly performing Gaussian blur and downsampling on feature map I2 to reflect the global low-frequency features. After obtaining the feature map, the feature fusion process is as follows: (1) Downsample feature map I1 to the same scale as feature map I2, denoted as feature map I1 ′ And input it into the ResNetBlock module together with the feature map I2 to complete feature fusion and enhancement, and obtain the feature map I′2; (2) Downsample the feature map I′2 to the same scale as the feature map I3, record it as feature map I″2, and input it into the ResNetBlock module together with the feature map I3 to obtain the fused enhanced feature map I′3; Finally, in the Laplacian reconstruction structure, the I′3, I′2, and I1 feature maps are upsampled layer by layer to restore the image to the original resolution, generating the preliminary output enhanced_features; then, the enhanced_features is residually added to the low-light image I to retain the global brightness information of the original low-light image I, and the final output I is obtained. ′ ; The Laplace multi-scale feature enhancement network structure LaMFE is used to enhance the background of low-light images, improve their brightness and contrast, while keeping the target area unchanged and avoiding feature interference; Based on the output result of rough detection D i = {B i ,C i ,S i }, accurately divide the target area and background area in the low-light image I; specifically, through the target detection frame B i The coordinates (x i ,y i ,w i ,h i ), the low-light image I is divided into a target area and a background area; the target area consists of the part covered by the detection frame, and the background area is the image part outside the target area; it is expressed as: Among them, R 目标 Indicates the target area, R(B i ) is the target detection box B i The corresponding area, R(I) is the pixel set of the low-light image I, R 背景 is the remaining background area; The division process is completed through coordinate calculation and image segmentation technology. First, the pixel range of each target area is accurately located according to the boundary coordinates of the target detection frame. Then, these target areas are subtracted from the low-light image I through logical operations to obtain the background area. Through the division of the target area and the background area, the brightness of the background area is adjusted in a targeted manner while keeping the characteristics of the target area unchanged. Step 4: Send the enhanced low-light image to the target detection model in step 2 again for processing, and observe the target detection effect after the background area is enhanced; through the detection confidence feedback, evaluate whether the current enhancement has an optimization effect on target detection; Assume that the image after background enhancement is I ′ , the process of coarse target detection algorithm is expressed as: D ′ =M(I ′ ) Among them, D ′ Represents the new detection result, including the updated target detection box B i ′ , target category label C i ′ and confidence score S i ′ ; By comparing D ′ and D, evaluate the effectiveness of background area enhancement; if the target detection box B i ′ More accurately covers the target area, while the confidence score S i ′ If the confidence score S is improved, it indicates that the current background area enhancement strategy is effective; otherwise, i ′ If the background image drops or the detection frame shifts, you may need to adjust the intensity or method of background enhancement. The enhancement effect is evaluated by the following indicators: ΔS i =S′ i -S i If ΔS i >0, background area enhancement is effective; otherwise, adjustment is required; Step 5: Repeat the loop process of background area enhancement and coarse target detection for n iterations in total; in each iteration, dynamically adjust the intensity of background area enhancement according to the detection results of the previous round; through multiple rounds of optimization, gradually improve the image quality and detection performance until the detection requirements are met; Specifically: By comparing the rough detection result D of the kth round (k) , according to the target detection box Target category label and confidence score changes, judging whether it is necessary to optimize the background area enhancement parameters; Specific strategies for adjusting background area enhancement include: 1) Increase or decrease the strength of the reinforcement: For the kth round of detection, compare the target confidence score of the previous round and the current confidence score change If the confidence score Increase that is ΔS i >0, indicating that the accuracy of target detection has improved, and the current background area enhancement strength may be effective, so the current enhancement strategy is maintained or strengthened; If the confidence score increases: maintain the current background area enhancement strength, that is, maintain the brightness increase factor α (k) constant; If the confidence score decreases, it means that the current background area is too enhanced, which may have excessively affected the characteristics of the target, thereby reducing the detection effect. In this case, reduce the brightness enhancement factor α (k) ;The adjustment formula of the brightness enhancement factor is: Wherein, Δα is the increment of the brightness factor, which is a preset constant or a dynamically calculated value, and is adjusted according to the magnitude of the confidence change; 2) Optimize and enhance the area: If the target detection box If there is a deviation, the target area R is re-divided accurately. 目标 and background area R 背景 , avoiding interference with the target area, while strengthening the optimization of the background area of ​​low-light images; During multiple rounds of iterations, the background enhanced image I ′(k+1) And the test results D (k+1) All depend on the output results of the previous round: Step 6: After completing n iterations, the optimized image is input into the precise target detection model for final detection; processing is performed on the enhanced image to obtain more accurate and stable target detection results; The final enhanced image I ′(n) Input a high-precision target detection algorithm M final Perform final testing to obtain accurate test results.

2. The multi-scale adaptive low-light target detection method based on intensity and region specificity according to claim 1, characterized in that: High-precision target detection algorithm M final Choose Faster R-CNN or RetinaNet; First, for the input image I ′(n) Perform feature extraction and target analysis, and process multi-scale targets through convolution feature pyramid; secondly, further adjust the position and size of the detection frame, and recalculate the target confidence of the target category; the detection process is described as: I ′(n) Input high-precision target detection algorithm M final , output the final target detection box D final , including target detection box, target category label and confidence score.

3. The multi-scale adaptive low-light environment target detection method based on intensity and region specificity according to claim 1, characterized in that: Final test result D final Contains all detected targets in the background area of ​​the low-light image; to ensure the detection effect, the comprehensive performance of the enhanced model LaMFE and the precise target detection model is verified by comparing with the standard detection results or manual annotations.

4. The multi-scale adaptive low-light environment target detection method based on intensity and region specificity according to claim 1, characterized in that: The low-light image has an average brightness value lower than 30, and a pixel value range of 0-255.

5. The multi-scale adaptive low-light environment target detection method based on intensity and region specificity according to claim 1, characterized in that: In step 2, the target detection model is YOLOv3.

6. The multi-scale adaptive low-light environment target detection method based on intensity and region specificity according to claim 1, characterized in that: In step 6, the precise target detection model is YOLOv8 or TOOD.

Images