A Multi-Scale Adaptive Target Detection Method Based on Intensity and Region Specificity in Low-Light Environments

Abstract

The present invention belongs to the field of object detection in deep learning networks, and discloses a multi-scale adaptive object detection method in low-light environments based on intensity and region specificity. By designing a Laplacian multi-scale feature enhancement network structure, the present invention significantly improves the object detection performance in low-light environments. LaMFE conducts meticulous optimization of low-light images by fusing multi-scale features and making full use of information at different scales. Meanwhile, an efficient residual connection mechanism is adopted to improve the efficiency and accuracy of feature extraction. This design not only improves the visualization quality of low-light images, but also significantly enhances the robustness and accuracy of the object detection model in complex low-light environments, making it perform outstandingly in a wide range of application scenarios.

Description

Technical Field

[0001] The present invention belongs to the field of object detection in deep learning networks, and particularly relates to an object detection method based on multi-scale flexible adaptability for low light intensity. Background Art

[0002] Object detection is a key task in the field of computer vision and plays a crucial role in a wide range of applications such as surveillance, autonomous driving, and military operations. The ability to accurately identify and locate objects in images is essential for the functionality and safety of these systems. However, a major 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, for example, night security cameras usually operate under limited ambient light, or in autonomous driving scenarios at dusk or even at night, where the surrounding scenes are dimly lit. In addition, military operations often occur under low visibility conditions, further emphasizing the necessity of researching powerful low-light object detection technologies.

[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 methods based on model optimization. Although traditional methods have improved the image brightness and quality to a certain extent, methods based on distribution mapping often ignore the illumination requirements of low-light images and the internal connections between pixels, which may lead to color distortion and abnormal details. Methods based on model optimization have limitations in prior design and are prone to color shift, underexposure, ghosting, and noise problems.

[0005] Deep learning-based low-light image enhancement methods originated in 2017 and have now become the mainstream methods for low-light image enhancement. Generally speaking, from the perspective of the achieved objectives, deep learning-based low-light image enhancement methods can be roughly divided into two categories: methods for brightness enhancement and methods for joint brightness enhancement and noise removal. The first category of methods improves the brightness of the image to display more detailed information of the image. The second category is the existing mainstream work, which combines brightness enhancement and noise removal and solves the brightness and noise problems in low-light images through a deep learning network, and can comprehensively improve the brightness and visual quality of the image.

[0006] Current low-light image enhancement technologies have evolved from traditional model design to data-driven deep learning. In terms of the learning mechanism, they are gradually moving towards semi-supervised / unsupervised learning and are no longer limited to fully supervised learning. At the same time, the evaluation method is also changing, from only focusing on visual quality to paying more attention to 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. The object detection methods under low-light conditions are mainly divided into two types: one is to first perform low-light image enhancement and then perform object detection on the enhanced image; the other is to directly optimize the enhancement network with the goal of object detection without generating the enhanced image.

[0008] The first method is usually called enhancement followed by detection. It first enhances the image under low-light conditions to improve the image quality and visibility, and then sends 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 specifically optimize the enhancement network to achieve better object detection performance. Different from traditional image enhancement algorithms, the optimization goal of this method is to minimize the detection loss, enabling object detection even without generating the enhanced image. 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 during the enhancement process.

[0010] Recently, there have been a series of advancements in low-light object detection that directly optimize the enhancement network with the goal of 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). It jointly learns CNN-PP and YOLOv3 to ensure that CNN-PP can learn an appropriate DIP in a weakly supervised manner to enhance the image for detection. MAET proposed by Cui et al. learns the intrinsic visual structure in a self-supervised manner by encoding and decoding the realistic illumination degradation transformation, while considering the physical noise model and image signal processing (ISP). Based on this representation, the decoded bounding box coordinates and classes are used to achieve the object detection task. To avoid excessive entanglement of the two tasks, the orthogonal tangent rule is applied to decouple the object and degradation features in MAET. This forms a parameterized manifold on which multi-task prediction can be geometrically formulated by maximizing the orthogonality between the tangents of the outputs of each task.

[0011] Multi-scale feature extraction and fusion is a key image processing technique aimed at extracting features from different scales and effectively fusing 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, adapting to targets of different sizes, shapes, and positions. In addition, multi-scale feature fusion can make full use of the complementarity of different-scale features, combining the detailed expressions of low-level features with the semantic information of high-level features to generate more robust feature representations. This technique is widely used in tasks such as object detection and image segmentation, significantly improving the adaptability of algorithms to complex scenes. In low-light object 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 object localization and recognition. Related research shows that multi-scale techniques can effectively improve low-light detection performance. For example, the DENet proposed by Qin et al. uses a Laplacian pyramid to decompose the input image into different frequency bands, enhancing the global brightness of the low-frequency component and the detailed information of the high-frequency component respectively, improving the expressiveness of low-light images. The FeatEnHancer module proposed by Hashmi et al. combines a multi-head attention mechanism, guiding the fusion of multi-scale features through task-related losses, improving the quality of feature representation. 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 propose a new solution to improve the performance of low-light object detection. We design a plug-and-play module - the Laplacian multi-scale feature enhancement (LaMFE) network structure. LaMFE can adaptively enhance the brightness of the background area locally with the overall intensity of the image, while limiting the noise interference introduced by the enhancement to the target area, thus significantly improving the object detection performance. In addition, common enhancement methods have different effects on different degrees of low-light environments: in cases where the light is not particularly dim, the enhancement method may be ineffective or even have side effects; while in extremely low-light conditions, only some enhancement algorithms are effective. To address this problem, our method, a multi-scale adaptive low-light environment object detection method based on intensity and region specificity, can flexibly select the best solution to achieve flexible low-light object detection. Summary of the Invention

[0013] In view of the above problems, we propose a Laplacian multi-scale feature enhancement (LaMFE) network structure, which significantly improves the object detection performance in low-light environments through multi-scale feature fusion and efficient residual connections. As a plug-and-play module, LaMFE can be directly applied to low-light images, locally enhance the background part, and support multiple loop processes until the expected effect is achieved. This flexibility enables LaMFE to greatly improve the accuracy and reliability of low-light object detection in various practical applications.

[0014] Technical solution of the present invention:

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

[0016] Step 1: Obtain a low-light image, which can be sourced from image acquisition devices such as surveillance cameras and night vision devices in low-light environments; use the low-light image as the initial input for target detection and background enhancement, and perform image enhancement, feature extraction, and target localization processing to improve the target detection performance under low-light conditions;

[0017] Step 2: Input the processed low-light image into a preselected target detection model, and use the feature extraction module in the target detection model to extract multi-scale features; through the analysis of the multi-scale features, locate potential target objects in the processed low-light image, and output the detection results, including the bounding box, class label, and confidence score (0-1) of the target. The confidence score is used to represent the reliability of the target detection. The preliminary target detection provides a rough position reference for the subsequent background enhancement step.

[0018] Perform a rough detection on the input low-light image I, and use the target detection model M to extract the preliminary information of the low-light image I. The specific process is expressed as D = M(I), where D is the detection result, including the target detection box B, the target class 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 position information of the target detection box, C i represents the target class label, and S i is the confidence score of the target detection model for the detection result, ranging from 0 to 1;

[0019] Through the rough detection, we can quickly locate the target area in the low-light image and output the rough position and class information of the target. The detection box B provides the spatial distribution of the target, the target class label C clarifies the nature of the target, and the confidence score S reflects the reliability of the detection. These information provide an important basis for the subsequent background enhancement, ensuring that the target area is not disturbed during the enhancement process, while optimizing the brightness characteristics of the background. The rough detection is not only a key step for preliminary target screening, but also provides a decision-making basis for the multi-round optimization of the system, helping to adjust the background enhancement intensity to gradually improve the target detection performance.

[0020] Step 3: According to the rough detection results, determine the target region in the low-light image I, and define the part around it as the background region; precisely divide the target region and the background region through coordinate calculation and image segmentation techniques, and only perform enhancement processing on the background region; construct a Laplacian multi-scale feature enhancement network structure LaMFE, including a Laplacian decomposition structure, a Laplacian reconstruction structure, and a ResNetBlock module for feature extraction; among them, the ResNetBlock module sequentially executes two Block modules to extract features. Assuming the input feature is x, the Block module sequentially performs group normalization, Swish activation function, Dropout, and a 3×3 convolutional layer on the feature x to obtain the feature x′, and then adds the feature x′ to the input feature x residually; in the Laplacian decomposition structure, set the decomposition level m to 3, and decompose the low-light image I through the Laplacian pyramid to convert it into three feature maps I1, I2, and I3 of different scales. The decomposition formula is as follows:

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

[0022] where 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: the feature map I1 is the reconstructed version after subtracting the blur from the low-light image I, retaining the high-frequency detail information in the low-light image I, including texture and edges; the feature map I2 is obtained by performing the same processing on the feature map I1 to extract lower-scale detail features; the feature map I3 is obtained by directly performing Gaussian blur and downsampling on the feature map I2, reflecting the global low-frequency features. After obtaining the feature maps, the feature fusion process is as follows:

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

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

[0025] Finally, in the Laplacian reconstruction structure, the feature maps I′3, I′2, and I1 are upsampled layer by layer to restore the image to the original resolution, generating the preliminary output enhanced_features. Subsequently, enhanced_features is added to the low-light image I residually to retain the global brightness information of the original image, obtaining the final output I′. 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, improving its brightness and contrast while keeping the target area unchanged and avoiding feature interference;

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

[0027] R 背景 = R(I)\R 目标

[0028] where R 目标 represents the target area, R(B i ) is the area corresponding to the target detection box B i , R(I) is the pixel set of the low-light image I, and 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 box; then, these target areas are subtracted from the low-light image I through logical operations to obtain the background area; this processing provides a clear area range for the subsequent background enhancement operation, enabling the enhancement process to focus on the background part without affecting the feature integrity of the target area. By dividing the target area and the background area, the brightness of the background area is adjusted specifically 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: Feed the enhanced low-light image back into the object detection model in Step 2 for processing, and observe the object detection effect after the background area is enhanced; through the detection confidence feedback, evaluate whether the current enhancement has optimized the object detection effect;

[0031] Assume the image after background enhancement is I′, and the process of the rough object detection algorithm is expressed as:

[0032] D′ = M(I′)

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

[0034] By comparing D′ and D, evaluate the effectiveness of the background area enhancement; if the object detection box B′ i more accurately covers the target area, and at the same time the confidence score S′ i increases, it indicates that the current background area enhancement strategy is effective; conversely, if the confidence score S′ i decreases or the detection box shifts, it may be necessary to adjust the intensity or method of background enhancement; evaluate the enhancement effect through the following indicators:

[0035] ΔS i = S′ i - S i

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

[0037] The purpose of repeating the rough detection is to gradually optimize the visibility and detection performance of the object in multiple rounds of enhancement and detection. The feedback of each round of rough detection will affect the intensity or range of the next round of background enhancement. For example, in some low-light scenarios, it may be necessary to increase the background brightness enhancement factor, while in other scenarios, it may be necessary to more precisely control the boundary of the enhanced 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 object detection performance, and lay a solid foundation for the final precise detection.

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

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

[0041] The specific strategies for adjusting the background area enhancement include:

[0042] 1) Increase or decrease the enhancement intensity:

[0043] For the k-th round of detection, compare the target confidence score of the previous round and the change in the current confidence score If the confidence score increases, i.e., ΔS i > 0, it indicates that the accuracy of target detection has improved, and the current background area enhancement intensity may be effective. Therefore, maintain or strengthen the current enhancement strategy;

[0044] If the confidence score rises: Keep the current background area enhancement intensity, i.e., maintain the brightness enhancement factor α (k) unchanged;

[0045] If the confidence score drops: It means that the current background area enhancement is too strong and may have overly affected the features of the target, thus reducing the detection effect; at this time, reduce the brightness enhancement factor α (k) ; The adjustment formula for the brightness enhancement factor is:

[0046]

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

[0048] 2) Optimize the enhancement area:

[0049] If the target detection box shows an offset, re-precisely divide the target area R 目标 and the background area R 背景 , avoid interfering with the target area, and at the same time strengthen the optimization of the background area of the low-light image;

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

[0051]

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

[0053] Step 6: After completing n iterations, input the optimized image into an accurate object detection model for final detection; process the enhanced image to obtain more accurate and stable object detection results; use a high-performance model to process the enhanced image to obtain accurate and stable object detection results suitable for practical application scenarios.

[0054] Input the finally enhanced image I′ (n) into a high-precision object detection algorithm M final for final detection to obtain accurate detection 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 has been significantly improved. At this time, input the finally enhanced image I′ (n) into a high-precision object detection algorithm M final for final detection to obtain accurate detection results. Compared with rough detection, the precise detection algorithm can better capture object details and improve the positioning accuracy and classification confidence.

[0056] Furthermore, the high-precision object detection algorithm M final selects Faster R-CNN or RetinaNet, and these models have the ability of multi-stage optimization; first, perform feature extraction and object analysis on the input image I′ (n) and process multi-scale objects through a convolutional feature pyramid; second, further adjust the position and size of the detection box and recalculate the object confidence of the target category; the detection process is described as: input I′ (n) into the high-precision object detection algorithm M final , and output the final object detection box D final , including the object detection box, object category label, and confidence score.

[0057] The final detection result D finalIt includes all detection targets in low-light environments, with relatively 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 object detection algorithms (YOLOv3 and TOOD) is verified by comparing with standard detection results or manual annotations. This stage makes full use of the enhanced image characteristics and the capabilities of high-precision detection algorithms, significantly improving the object detection performance in low-light environments and meeting the requirements for efficient and highly reliable detection in practical applications.

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

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

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

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

[0062] The beneficial effects of the present invention: By designing the Laplacian multi-scale feature enhancement (LaMFE) network structure, the present invention significantly improves the object detection performance in low-light environments. LaMFE conducts meticulous optimization of low-light images by fully utilizing information at different scales through multi-scale feature fusion, and at the same time adopts 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 enhances the robustness and accuracy of the object detection model in complex low-light environments, making it perform excellently in a wide range of application scenarios.

[0063] The core advantage of the LaMFE module lies in its combination with rough detection to form a multi-round cyclic enhancement mechanism. While focusing on local optimization of the background area, it keeps the detection targets unchanged. During the cyclic process, the rough detection module provides position and category information by initially identifying the target area, providing clear guidance for background enhancement, enabling each round of enhancement to focus on improving the background brightness and contrast. At the same time, the rough detection can also evaluate the enhancement effect and dynamically adjust the intensity and range of background enhancement according to the detection results to ensure that the optimization process converges gradually. This effectively avoids the defect of traditional methods being prone to overprocessing, not only ensuring the integrity of the detection targets but also significantly reducing the side effects that may be caused by enhancement operations.

[0064] In addition, LaMFE adopts a plug-and-play modular design and can be flexibly integrated into existing object detection frameworks. By leveraging the synergy between rough detection and multi-round enhancement, LaMFE achieves high adaptability, enabling it to dynamically adjust the enhancement strategy and the number of cycles according to the specific scenario, thus ensuring that the final enhancement effect meets expectations. This flexibility and reliability make LaMFE exhibit significant application value in image processing and object detection tasks in low-light environments, providing an efficient and stable solution for related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a detailed training framework structure diagram of the object detection method based on intensity and region adaption in low-light environments provided by the present invention;

[0066] Figure 2 It is a flowchart of the object detection method based on intensity and region adaption in low-light environments provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] The following further describes the specific embodiments of the present invention in conjunction with the drawings and technical solutions.

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

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

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

[0071] First, obtain pairs of low-light image training data. The images can be obtained by taking and annotating images under low-light conditions by oneself, or using publicly available low-light datasets such as Exdark or DarkFace, etc. Then, crop the image size to 408 pixels for the short side and 1333 pixels for the long side to ensure that the input image sizes are consistent. Then, perform a series of data augmentation operations, including random cropping (randomly select a region from the original image for cropping), random rotation (randomly rotate the image by a certain angle, such as between -10 degrees and +10 degrees), random flipping (flip the image horizontally or vertically), and random adjustment of brightness and contrast, etc. Through these augmentation operations, the robustness of the model can be improved, overfitting can be avoided, and its object detection performance in low-light environments can be enhanced.

[0072] 2. Preliminary object detection

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

[0074] The model M outputs preliminary detection results D. For example, assume the model detects two persons, namely D1 and D2, where:

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

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

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

[0078] The final detection result is:

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

[0080] This indicates that the model has successfully detected two objects, which are two objects of the "person" category, and output the detection box, class label, and confidence score of each object.

[0081] 3. Background Region Determination and Enhancement

[0082] Next, the object 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 starting from the image coordinates, with a width of 100 and a height of 200. This area contains a person and is the target region R 目标 . The part outside the target region is defined as the background region.

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

[0084] Next, we use the LaMFE model to enhance the background area R 背景 . Suppose after LaMFE enhancement, the brightness of the background is increased by 30% and the contrast is increased by 25%, making the background part more clearly visible. At the same time, the target area R 目标 remains unchanged to avoid any feature interference.

[0085] In the finally 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. Target Detection and Feedback after Enhancement

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

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

[0089] D′ = M(I′)

[0090] Suppose through detection, the new target detection box obtained is B′1 = {x′1 = 48, y′1 = 58, w′1 = 102, h′1 = 202}, the target class 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 increased.

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

[0092] ΔS1 = S′1 - S1 = 0.92 - 0.85 = 0.07

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

[0094] 5. Multi-round Iterative Optimization

[0095] The loop process of repeatedly performing background region enhancement and rough target detection is carried out for 3 iterations; in the first iteration, the image after background enhancement is I′1, and the detection result output by the target detection model is Compared with the detection result of the first round, the confidence score is higher than the initial by 0.07. Therefore, This indicates that the background enhancement strategy is effective, so we decide to continue to maintain the current enhancement strategy in the second iteration.

[0096] In the second 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 time, the position of the target box remains stable, and the confidence score is higher than the previous round by 0.01. After calculation, Since it is considered that the current background enhancement intensity is still effective, so the brightness enhancement factor α (2) remains unchanged.

[0097] In the third iteration, the enhanced image I′2 is used as the input for detection again, and the target detection result is At this time, the change in the position information of the target box is small, and the confidence score is higher than the previous round by 0.01. It is concluded that At this time, since we continue to maintain the current background enhancement strategy, and the brightness enhancement factor α (3) also remains unchanged. After multiple iterations, after 3 optimizations, the position and size of the target box gradually stabilize, and the confidence score gradually increases, from the initial to an increase of about 10.6%. During this process, the enhancement of the background region effectively improves the brightness and contrast of the image while keeping the features of the target region 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 region enhancement and target detection optimization, the obtained image I′ (3) is sent to the high-precision target detection model YOLOv8 for final detection. After processing, the target detection result is:

[0100]

[0101] Among them, the target box precisely locates the target position, and the confidence score is 0.95, indicating a high target detection accuracy.

[0102] The present invention improves the target detection accuracy in low-light images through precise background region enhancement and target detection optimization strategies. After preliminary target detection, the target region and the background region are determined based on the target detection box, and the LaMFE model is used to enhance the background to make the background brighter and clearer while keeping the target region unchanged to avoid interference. Through multiple rounds of iterative optimization, the enhancement intensity is dynamically adjusted according to the detection results of each round to continuously improve the detection confidence and the accuracy of the target box. 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, significantly improving the performance of target detection in low-light environments.

Claims

1. A multi-scale adaptive target detection method based on intensity and region specificity in low-light environments, characterized in that, It includes the following steps: Step 1: Obtain a low-light image, which is used as the initial input for object detection and background enhancement. Through image enhancement, feature extraction, and object localization processing, the object detection performance under low-light conditions is improved; Step 2: Input the processed low-light image into a preselected object detection model. Use the feature extraction module in the object detection model to extract multi-scale features; through the analysis of multi-scale features, potential target objects are located in the processed low-light image, and the detection results are output, including the bounding box, class label, and confidence score of the target, providing basic data for subsequent tasks; Perform a rough detection on the input low-light image I, and use the object detection model M to extract the preliminary information of the low-light image I. The specific process is expressed as D = M(I), where D is the detection result, including the object detection box B, the object 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 position information of the object detection box, C i represents the object category label, and S i is the confidence score of the object detection model for the detection result, ranging from 0 to 1; Step 3: According to the rough detection results, determine the target region in the low-light image I, and define the part around it as the background region; precisely divide the target region and the background region through coordinate calculation and image segmentation techniques, and only perform enhancement processing on the background region; construct a Laplacian multi-scale feature enhancement network structure LaMFE, including a Laplacian decomposition structure, a Laplacian reconstruction structure, and a ResNetBlock module for feature extraction; among them, the ResNetBlock module sequentially executes two Block modules to extract features. Assuming the input feature is x, the Block module sequentially performs group normalization, Swish activation function, Dropout, and a 3×3 convolutional layer on the feature x to obtain the feature x′, and then adds the feature x′ to the input feature x residually; in the Laplacian decomposition structure, set the decomposition level m to 3, and decompose the low-light image I through the Laplacian pyramid to convert it into three feature maps I1, I2, and I3 at different scales. The decomposition formula is as follows: I1 = I - U(D(G(I))), I2 = I1 - U(D(G(I1))), I3 = D(G(I2)) where G represents the Gaussian blur process, D represents downsampling, and U represents upsampling; after obtaining the feature maps, the feature fusion process is as follows: (1) Downsample the feature map I1 to the same scale as the feature map I2, denoted as the feature map I1 ′ , and input it into the ResNetBlock module together with the feature map I2 to complete feature fusion and enhancement, obtaining the feature map I2 ′ ; (2) Downsample the feature map I2 ′ to the same scale as the feature map I3, denoted as the feature map I2 ″ , and input it together with the feature map I3 into the ResNetBlock module to obtain the fused and enhanced feature map I3 ′ ; Finally, in the Laplacian reconstruction structure, I3 is upsampled layer by layer ′ , I2 ′ and the feature maps of I1 are upsampled to restore the images to the original resolution, generating preliminary output enhanced features; subsequently, the enhanced features are added to the low-light image I residually to retain the global brightness information of the original low-light image I, obtaining the final output I ′ ; the Laplacian multi-scale feature enhancement network structure LaMFE is used to enhance the background of the low-light image, improving its brightness and contrast while keeping the target area unchanged to avoid feature interference; Based on the output result D of the rough detection i ={B i , C i , S i}, the target region and the background region in the low-light image I are accurately divided; specifically, through the coordinates (x i , y i , w i , h i ) of the target detection box B i ), the low-light image I is divided into a target region and a background region; the target region consists of the part covered by the detection box, and the background region is the part of the image outside the target region; Step 4: Send the enhanced low-light image into the object detection model in Step 2 again for processing, and observe the object detection effect after the background region is enhanced; through the detection confidence feedback, evaluate whether the current enhancement has optimized the object detection effect; Step 5: Repeat the loop processing of background region enhancement and rough object detection for a total of n iterations; in each iteration, dynamically adjust the intensity of background region enhancement according to the detection results of the previous round by adjusting the brightness enhancement factor; through multiple rounds of optimization, gradually improve the image quality and detection performance until the detection requirements are met; Step 6: After completing n iterations, input the optimized image into the precise object detection model for final detection; perform processing on the enhanced image to obtain more accurate and stable object detection results; Input the finally enhanced image I ′(n) into a high-precision object detection algorithm M final for final detection to obtain accurate detection results.

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

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

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

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

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

7. The multi-scale adaptive target detection method in low-light environments based on intensity and region specificity according to claim 1, wherein In step 3, the three decomposed feature maps at different scales cover feature information at different scales: the feature map I1 is a reconstructed version obtained by subtracting the blur from the low-light image I, retaining the high-frequency detail information in the low-light image I, including textures and edges; the feature map I2 is obtained by performing the same processing on the feature map I1 to extract lower-scale detail features; the feature map I3 is obtained by directly performing Gaussian blur and downsampling on the feature map I2, reflecting the global low-frequency features; [[ID= Among them, R 目标 represents the target region, and R(B i ) is the region corresponding to the target detection box B i . R(I) is the set of pixels of the low-light image I, and R 背景 is the remaining background region; ​ 8. The multi-scale adaptive target detection method in low light environment based on intensity and region specificity according to claim 1, characterized in that Step 4 is specifically as follows: The image after background enhancement is I ′ , and the process of the rough 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 ′ , the target class label C i ′ and the confidence score S i ′ ; By comparing D ′ and D, evaluate the effectiveness of background area enhancement; if the target detection box B i ′ covers the target area more accurately, while the confidence score S i ′ increases, it indicates that the current background area enhancement strategy is effective; conversely, if the confidence score S i ′ decreases or the detection box shifts, it may be necessary to adjust the intensity or method of background enhancement; evaluate the enhancement effect through the following metrics: ΔS i = S i ′ - S i If ΔS i > 0, the background area enhancement is effective; otherwise, adjustment is required.

9. The multi-scale adaptive target detection method in low light environment based on intensity and region specificity according to claim 1, characterized in that, Step 5 is specifically as follows: By comparing the rough detection results D in the k-th round (k) , according to the target detection box target category label and the confidence score changes, determine whether it is necessary to optimize the background region enhancement parameters; ​ ​ For the k-th round of detection, compare the target confidence score of the previous round with the change in the current confidence score If the confidence score increases, i.e., ΔS i > 0, it indicates that the accuracy of target detection has improved, and the current enhancement intensity of the background area may be effective. Therefore, maintain or strengthen the current enhancement strategy; If the confidence score increases: Keep the current background area enhancement intensity, i.e., maintain the brightness enhancement factor α (k) unchanged; If the confidence score drops: It indicates that the current background area enhancement is too strong, which may have overly affected the features of the target, thus reducing the detection effect; at this time, reduce the brightness enhancement factor α (k) ; The adjustment formula for the brightness enhancement factor is: ​ ​ If the target detection box shows an offset, re-precisely divide the target region R 目标 and the background region R 背景 , avoid interference with the target region, and at the same time strengthen the optimization of the background region of the low-light image; In the multi-round iteration process, the background-enhanced image I of each round ′(k+1) and the detection result D (k+1) both depend on the output result of the previous round: 。

Images