Systematized semantic enhancement method and device for low-illumination video image change detection

Abstract

The invention discloses a semantic enhancement method and device for low-illumination video image change detection, and the method comprises the steps: a time sequence semantic memory mechanism module carries out the modeling of cross-frame long-term semantic consistency based on a dual-temporal image through bidirectional time sequence interaction and context semantic enhancement, and outputs a feature representation with long-term semantic consistency; performing multi-scale processing on the feature representation of the long-term semantic consistency to obtain multi-scale fusion features, dynamically integrating the multi-scale fusion features by a progressive multi-scale semantic fusion network module through adaptive weighting, semantic guidance and low-illumination enhancement strategies, and outputting multi-scale fusion feature representation; and obtaining a final change detection result according to the multi-scale fusion feature representation. The device comprises a memory and a processor. Through the systematic semantic enhancement design, the noise interference is effectively relieved, the cross-frame semantic consistency is kept, and the small target detection capability under the low illumination condition is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image change detection, in particular to a systematic semantic enhancement method and device for low-light video image change detection, and the present application is applicable to automatic monitoring, traffic monitoring and environmental analysis under complex lighting conditions, and uses deep learning technology to improve the detection capability of small changes and small targets in video sequences. BACKGROUND

[0002] Low-light video image change detection is crucial for night city monitoring, and can identify subtle target movements under insufficient lighting conditions. However, there are problems such as poor visibility, low contrast, complex time sequence dynamics, and lack of semantic information for small targets. In order to solve these problems, researchers have developed various deep learning-based methods for modeling spatial and temporal relationships in the change detection task. Common methods include using convolutional neural networks (CNN) for feature extraction, using attention mechanisms to focus on key areas, and capturing inter-frame changes through temporal modeling. Currently, research and technology for low-light large field of view scenes show obvious limitations, especially in the context of change detection. Existing methods (including traditional image difference, optical flow technology and early machine learning methods) are difficult to effectively perform under weak light conditions. The main defect is the inability to accurately distinguish between real changes (such as the movement or structural changes of pedestrians and vehicles) and pseudo-changes caused by uneven lighting, shadows or noise.

[0003] Although existing methods have made progress, there are still the following limitations: most methods are designed for well-lit or static images and are difficult to handle high noise problems in low-light videos; their temporal modeling often cannot maintain the long-term semantic consistency required for small targets; in addition, the multi-scale feature fusion capability is weak and the boundary perception is not strong, which limits the detection accuracy of small targets in low-contrast environments. More importantly, there is currently a lack of research on semantic enhancement for low-light change detection, and there is a lack of a systematic semantic enhancement framework specifically for low-light video image change detection. SUMMARY

[0004] The present application provides a systematic semantic enhancement method and device for low-light video image change detection, which effectively alleviates noise interference, maintains cross-frame semantic consistency, and significantly improves the detection capability of small targets under low-light conditions, as described in detail below:

[0005] In a first aspect, a semantic enhancement method for low-light video image change detection is provided, the method comprising:

[0006] The time sequence semantic memory mechanism module models the long-term semantic consistency across frames based on the dual-phase images through bidirectional time interaction and context semantic enhancement, and outputs feature representation with long-term semantic consistency.

[0007] The feature representation with long-term semantic consistency is subjected to multi-scale processing to obtain multi-scale fusion features, and the progressive multi-scale semantic fusion network module dynamically integrates the multi-scale fusion features through adaptive weighting, semantic guidance and low-light enhancement strategies, and outputs multi-scale fusion feature representation.

[0008] The final change detection result is obtained according to the multi-scale fusion feature representation.

[0009] The final change detection result is obtained according to the multi-scale fusion feature representation.

[0010] The edge-aware multi-scale change detection module detects changes in the video through multi-scale edge detection and adaptive scale attention mechanism based on the multi-scale fusion feature representation, ensuring the target detection effect in low-light environment.

[0011] The time sequence semantic memory mechanism module uses a bidirectional time interaction mechanism to perform forward enhancement calculation, and generates weights by Sigmoid activation function after concatenating and convolving the current time feature and the next time feature.

[0012] The original feature is multiplied element by element to obtain the forward enhanced feature; then the reverse semantic enhancement is performed, and similar operation process but the feature order is exchanged, and finally the features after forward and reverse enhancement are fused through concatenation and convolution operation to realize the integration of bidirectional time information.

[0013] The progressive multi-scale semantic fusion network module includes a semantic guidance module, which generates channel-level weight semantic attention module through high-level feature, combines spatial attention module of average pooling and maximum pooling for target positioning, and enhances the performance of low-scale features in low-light environment through progressive semantic propagation mechanism starting from the deepest feature in the network feature pyramid.

[0014] The progressive multi-scale semantic fusion network module further includes:

[0015] The semantic brightness adaptive module adjusts the brightness according to the intensity distribution of semantic features to enhance the semantic visibility of dark areas;

[0016] The semantic contrast enhancement module improves the discriminability of semantic features through multiplication operation to highlight the semantic features of small targets;

[0017] The semantic noise suppression module uses depth separable convolution to remove noise while preserving key semantic edge details.

[0018] The edge-aware multi-scale change detection module is configured to aggregate edge features by edge-aware difference calculation, and employ a Sobel operator and a Laplacian operator, multi-scale convolution kernels, and a hole convolution.

[0019] In a second aspect, a semantic enhancement device for low-light video image change detection is provided, and the device comprises a processor and a memory.

[0020] In a third aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores a computer program.

[0021] The technical scheme provided by the present application has the following beneficial effects:

[0022] 1. The present application first proposes a semantic enhancement framework for low-light video image change detection, and designs a TSMM (time sequence semantic memory mechanism) module to capture long-term semantic consistency across frames through bidirectional time interaction and context semantic enhancement, thereby achieving accurate detection of small target micro changes.

[0023] 2. The present application innovatively combines semantic enhancement with multi-scale processing, proposes a progressive multi-scale semantic fusion network method, dynamically fuses multi-scale features through adaptive weighting, semantic guidance, and low-light enhancement, and guarantees the semantic visibility of small targets in low-contrast scenes.

[0024] 3. The present application constructs an edge-aware semantic detection mechanism, proposes an EMCD (edge-aware multi-scale change detection module) module that combines multi-scale edge detection and scale attention mechanisms, and significantly improves the boundary segmentation accuracy and semantic accuracy of small targets in low-light videos. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of the systematic semantic enhancement method for low-light video image change detection is shown.

[0026] Figure 2 A block diagram of the systematic semantic enhancement for low-light video image change detection is shown.

[0027] Figure 3 A memory intention of the proposed TSMM is shown.

[0028] Figure 4 A schematic diagram of the proposed progressive multi-scale semantic fusion network (PMSF) is shown. DETAILED DESCRIPTION

[0029] To make the objects, technical solutions and advantages of the present application clearer, the embodiments of the present application are described in further detail below.

[0030] To overcome the above shortcomings, the embodiments of the present application first propose a semantic enhancement framework specially applicable to low-illumination video image change detection, which integrates three innovative modules: a time-series semantic memory mechanism module (TSMM) that models historical semantic consistency through bidirectional time-series interaction and context semantic enhancement, and enhances time-series robustness; a progressive multi-scale semantic fusion (PMSF) that dynamically fuses multi-scale features through adaptive semantic guidance and low-illumination enhancement mechanism, and improves image semantic visibility; and an edge-aware multi-scale change detection module (EMCD) that improves boundary accuracy and detection accuracy in combination with an edge-guided semantic detection strategy. The framework effectively alleviates noise interference, maintains cross-frame semantic consistency, and significantly improves the detection capability of small targets under low-illumination conditions through systematic semantic enhancement design.

[0031] Embodiment 1

[0032] A semantic enhancement method for low-illumination video image change detection, referring to Figure 1 and Figure 2 The method comprises the following steps:

[0033] S11: an input module for receiving a pair of low-light video images taken at different times as double-time images;

[0034] Among them, the received low-light video images (i.e. double-time images) are used for image processing and feature extraction in subsequent steps S12 to S14. Through the double-time memory network (TSMM) module, the input low-light video images will achieve consistency in time, thereby enhancing the image quality, which provides a better basis for subsequent analysis. In the multi-scale feature fusion network (PMSF), different levels of features are extracted and fused using the input images of step S11 to effectively enhance the performance of low-light videos. In addition, the multi-scale change detection module (EMCD) will rely on the multi-scale features extracted previously to detect changes in the video to ensure the target detection effect in low-light environments.

[0035] S12: a time-series semantic memory mechanism (TSMM) module that models the long-term semantic consistency across frames through bidirectional time-series interaction and context semantic enhancement based on the double-time images, and outputs a feature representation with long-term semantic consistency;

[0036] S13: Multi-scale fusion features are obtained by multi-scale processing of long-term semantically consistent feature representations. The Progressive Multi-Scale Semantic Fusion Network (PMSF) module dynamically integrates multi-scale fusion features through adaptive weighting, semantic guidance and low-light enhancement strategies, and outputs multi-scale fusion feature representations.

[0037] S14: The final change detection result is obtained based on the multi-scale fusion feature representation.

[0038] Among them, the edge-aware multi-scale change detection module (EMCD) relies on the multi-scale fusion feature representation extracted in step S13. Through multi-scale edge detection and adaptive scale attention mechanism, it enhances the boundary characterization ability, detects changes in the video, and ensures the target detection effect in low-light environment.

[0039] In summary, the embodiments of the present invention effectively mitigate noise interference, maintain cross-frame semantic consistency, and significantly improve the detection capability of small targets under low-light conditions through semantic enhancement design.

[0040] Example 2

[0041] The feasibility of the scheme in Example 1 is verified below, using specific experimental data and calculation formulas. See the description below for details:

[0042] S21: Construct low-light datasets, including: L2CD dataset and LL-PSCD dataset;

[0043] The L2CD dataset was self-collected in a real-world environment using high-resolution cameras with low-light sensitivity, filmed at night. The cameras were fixed in high positions on campus, such as in teaching buildings and libraries. The dataset primarily targets dynamic objects such as pedestrians, vehicles, and trees, reflecting typical security and traffic monitoring scenarios. Video was continuously acquired over a relatively long period to capture dynamic information changing over time, such as object movement and environmental changes. The raw video underwent a systematic processing workflow, extracting images at intervals of 10 to 50 frames to ensure temporal dynamic diversity. A total of 880 pairs of dual-temporal image pairs were generated, with an image resolution of 1280×720 pixels. Each processed image pair was manually annotated with a change mask, accurately depicting subtle changes in small targets such as pedestrian movement or slight changes in vegetation under low-light conditions.

[0044] The LL-PSCD dataset is an enhanced version derived from the publicly available PSCD (Panoramic Semantic Change Detection) dataset. This processed dataset contains 2300 pairs of dual-temporal images, finely processed to simulate low-light conditions, providing a valuable resource for testing change detection algorithms in low-light environments. The original PSCD dataset is renowned for its diverse pedestrian activities and environmental change scenarios. This embodiment of the invention transforms it through a controlled low-light preprocessing workflow, including: reducing brightness by 50%, setting the contrast adjustment factor to 0.3, and adding Gaussian noise with a mean of 0 and a standard deviation of 10 to simulate the limitations of sensors in low-light conditions in reality. All images are uniformly set to a resolution of 1280×720 pixels. S22: The TSMM module employs a bidirectional temporal interaction mechanism. First, it performs forward enhancement computation, concatenating and convolving the current time-step features with the next time-step features, then generating weights using the Sigmoid activation function. These weights are then multiplied element-wise with the original features to obtain the forward-enhanced features. Next, it performs reverse semantic enhancement, using a similar process but swapping the feature order. Finally, it fuses the forward and reverse-enhanced features through concatenation and convolution, achieving effective integration of bidirectional temporal information.

[0045] Forward semantic enhancement:

[0046]

[0047] Reverse semantic enhancement:

[0048]

[0049] Where F1 and F2 represent the features of the images before and after the event, This represents element-wise multiplication. This represents the Sigmoid activation function. This represents the convolution operation. .

[0050] The final bidirectional features are fused in the following way:

[0051]

[0052] in, Indicates forward semantic passing; This indicates backward semantic passing.

[0053] S23: As Figure 3 The TSMM module shown includes four-scale multi-scale memory units, which are processed using a simplified iterative multi-scale dense block (SiMDB) with a distillation rate of 0.25 to optimize semantic features and enhance the detection capability of small changes in low-light video frames.

[0054] S24: The TSMM module extracts global semantic context information through adaptive average pooling, combines multi-scale convolution operations to achieve local semantic awareness, and uses a connected attention weighting strategy to fuse the extracted features to ensure the integrity and consistency of the features.

[0055] S25: The PMSF module employs an adaptive weighting mechanism, dynamically adjusting the importance weights of features at different scales by calculating the quality score for each scale. The quality score calculation uses the Sigmoid activation function to perform global average pooling on the features. For example... Figure 4 The PMSF module shown includes an adaptive weighting module, which calculates the quality score for the i-th scale using the following formula:

[0056]

[0057] in, This represents the Sigmoid activation function. Indicates the feature map Perform global average pooling. Then concatenate the quality scores from all scales to generate a global weight vector.

[0058]

[0059] The weighted features at each scale are obtained using the following formula:

[0060]

[0061] in, Represents element-wise multiplication. Representing feature maps, This indicates that the adaptive weighting module is the feature map at the i-th scale. This represents the weighted features at each scale.

[0062] This method ensures that high semantic quality features associated with small targets receive higher weights, thereby increasing their contribution to the overall detection.

[0063] S26: The PMSF module includes: a semantic guidance module, which generates a semantic attention module with channel-level weights through high-level features (i.e., the 4th scale feature map), a spatial attention module that combines average pooling and max pooling for target localization, and enhances the performance of low-scale features in low-light environments through a progressive semantic propagation mechanism that starts from the deepest (topmost) features in the network feature pyramid.

[0064] S27: The EMCD module uses edge-aware differential computation, employing Sobel and Laplacian operators, multi-scale convolutional kernels (3×3 and 5×5), and dilated convolutions to aggregate edge features. The calculation formula is as follows:

[0065]

[0066] in, This represents the edge feature map extracted by the Sobel operator. This represents the edge feature map extracted by the Laplacian operator. This represents the geometric edge feature maps extracted by multi-scale convolutional kernels (3×3 and 5×5). This represents the macroscopic edge feature map extracted by dilated convolution.

[0067] The edge-guided differential features are represented as follows:

[0068]

[0069] in, This is the aggregated edge feature map. This is the activation function.

[0070] S28: The EMCD module constructs four parallel processing branches to process targets of different scales: the small target branch uses 1×1 to 3×3 convolutions, the medium target branch uses 3×3 to 5×5 convolutions, the large target branch uses 5×5 to 7×7 convolutions, and the super-large target branch uses dilated convolutions with dilation rates of 2 and 4. Adaptive feature fusion is achieved by calculating the attention weights of each scale through the Softmax function.

[0071] The formula for calculating the scale attention weight is as follows:

[0072]

[0073] The final feature representation is calculated as follows:

[0074]

[0075] in, Let J be the convolution weights of the j-th dilated convolution path in the s-th scale branch. This is the inter-frame difference feature map after edge enhancement. For normalization, Let be the global channel weight vector at the i-th scale. Let be the features of the i-th path.

[0076] Through the complete steps outlined above, accurate change detection in low-light environments is ultimately achieved. Specifically, the system receives a pair of low-light video images captured at different times as input. Key steps include establishing cross-frame semantic consistency through a temporal semantic memory mechanism, dynamically integrating multi-level feature information through a progressive multi-scale semantic fusion network, and strengthening boundary representation through an edge-aware multi-scale change detection module. Finally, a weighted fusion strategy generates an accurate change detection map. This detection map effectively distinguishes between changed and unchanged pixels, accurately marking the area of ​​change between the two temporal images. This effectively solves the technical challenge of low change detection accuracy in traditional methods under low-light conditions at night due to insufficient light, noise interference, and low contrast. The entire process not only maintains sensitivity to changes in small targets but also enhances robustness to complex lighting changes, providing reliable technical support for practical applications such as campus nighttime monitoring and urban security. It achieves an intelligent transformation from "blurry low-light image pairs" to "accurate change detection results."

[0077] Example 3

[0078] To evaluate the performance of the proposed network model, this embodiment uses Precision, Recall, F1, Overall Accuracy, and Intersection over Union as evaluation metrics. P: Recall is the proportion of correctly predicted changed pixels out of all pixels predicted as changed. R: Recall is the proportion of correctly predicted changed pixels out of the total number of truly changed pixels. F1 score: The harmonic mean of precision and recall, measuring the balance between detection performance. OA: The proportion of all pixels correctly classified (changed vs. unchanged). IoU: The ratio of the intersection to the union of predicted changed regions and truly changed regions, measuring the degree of spatial overlap.

[0079] To illustrate the accuracy of the method, comparisons are made with various models on low-light datasets across two datasets. Overall, the method of this invention performs excellently across all evaluation metrics, achieving the highest performance, specifically as shown in the results on the L2CD dataset:

[0080] As shown in Table 1, our method achieved excellent performance on the L2CD dataset: F1 score: 81.57%, precision: 90.71%, recall: 74.1%, IoU: 73.95%, and overall accuracy: 90.05%, demonstrating the best performance among all compared methods. Compared to the top-performing algorithm RCTNet, our method improves the F1 score by 8.25% and the IoU by 15.48%, thanks to the TSMM module's ability to extract temporal semantic relationships, which helps distinguish between real changes and false detections in complex environments. Compared to TinyCD, the F1 score is improved by 40.67%, precision by 23.84%, and IoU by 41.82%, highlighting the limitations of lightweight methods when dealing with scenarios requiring complex feature modeling. Compared to the Transformer-based BIT method, our method improves the F1 score by 21.58% and the IoU by 28.96%, indicating that our method outperforms the pure attention mechanism in a hybrid architecture based on temporal memory, multi-scale processing, and edge awareness.

[0081] Further analysis shows that this method significantly improves recall while maintaining high accuracy, mainly attributed to the adaptive weighting and semantic guidance of the PMSF module, which effectively alleviates the problem of missed detection of small targets. Compared with Transformer methods such as TransY-Net and TransCD, this method shows significant improvements in both accuracy and IoU, validating the advantages of structured temporal modeling and multi-scale feature enhancement.

[0082] On the more challenging LL-PSCD dataset, our method remains robust, achieving: F1 score of 79.06%, precision of 79.07%, recall of 79.05%, IoU of 65.37%, and overall accuracy of 89.64%. Compared to RCTNet, the F1 score is improved by 3.66%, and the IoU by 4.85%, indicating that the edge perception processing of the EMCD module and the low-light enhancement capabilities of the PMSF module are particularly important in complex scenes. Compared to TinyCD, the F1 score is improved by 6.13%, and the IoU by 8.65%, while maintaining a balance between precision and recall, demonstrating that our method effectively addresses the shortcomings of lightweight methods in complex environments. Compared to the BIT method, our method improves F1 by 13.37% and IoU by 16.46%, indicating that the TSMM module provides crucial contextual information that Transformer methods cannot capture under low-light conditions. Compared to traditional CNN methods such as SNUNet, this method achieves significant superiority in all metrics, with particularly remarkable improvements in recall and IoU, validating the effectiveness of the multi-module design of this method for small object detection.

[0083] Table 1

[0084]

[0085] Example 4

[0086] A semantic enhancement apparatus for detecting changes in low-light video images, the apparatus comprising: a processor and a memory, wherein the memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the apparatus to perform the following method steps in Embodiment 1:

[0087] The temporal semantic memory mechanism module is based on dual-temporal images. Through bidirectional temporal interaction and contextual semantic enhancement, it models long-term semantic consistency across frames and outputs feature representations with long-term semantic consistency.

[0088] Multi-scale fusion features are obtained by performing multi-scale processing on long-term semantically consistent feature representations. The progressive multi-scale semantic fusion network module dynamically integrates multi-scale fusion features through adaptive weighting, semantic guidance and low-light enhancement strategies, and outputs multi-scale fusion feature representations.

[0089] The final change detection result is obtained based on the multi-scale fusion feature representation.

[0090] The final change detection result obtained based on the multi-scale fusion feature representation is as follows:

[0091] The edge-aware multi-scale change detection module detects changes in videos based on multi-scale fusion feature representations, through multi-scale edge detection and an adaptive scale attention mechanism, ensuring target detection performance in low-light environments.

[0092] Among them, the temporal semantic memory mechanism module adopts a bidirectional temporal interaction mechanism to perform forward enhancement calculation, and then concatenates and convolves the current time feature with the next time feature and generates weights through the Sigmoid activation function.

[0093] The forward-enhanced features are obtained by multiplying the original features element-wise; then, the reverse semantic enhancement is performed, using a similar operation process but swapping the order of the features. Finally, the forward-enhanced and reverse-enhanced features are fused through concatenation and convolution operations to achieve the integration of bidirectional temporal information.

[0094] The progressive multi-scale semantic fusion network module includes a semantic guidance module, which generates a semantic attention module with channel-level weights through high-level features, combines a spatial attention module with average pooling and max pooling for target localization, and enhances the performance of low-scale features in low-light environments through a progressive semantic propagation mechanism starting from the deepest features in the network feature pyramid.

[0095] The progressive multi-scale semantic fusion network module further includes:

[0096] The semantic brightness adaptive module adjusts the brightness according to the intensity distribution of semantic features to enhance the semantic visibility of dark areas;

[0097] The semantic contrast enhancement module improves the discriminability of semantic features through multiplication operations, highlighting the semantic features of small targets;

[0098] The semantic noise suppression module uses depthwise separable convolution to remove noise while preserving key semantic edge details.

[0099] The edge-aware multi-scale change detection module aggregates edge features through edge-aware differential calculation, employing the Sobel and Laplacian operators, multi-scale convolution kernels, and dilated convolution. It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and will not be repeated here.

[0100] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0101] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0102] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0103] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0104] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0105] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0106] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer-readable storage medium can be any usable medium accessible to a computer or a data storage device such as a server or data center that integrates one or more usable media. The usable medium can be a magnetic medium or a semiconductor medium, etc. Unless otherwise specified, the model numbers of the devices in this embodiment of the invention are not limited; any device capable of performing the above functions is acceptable.

[0107] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semantic enhancement method for detecting changes in low-light video images, characterized in that, The method includes: The temporal semantic memory mechanism module is based on dual-temporal images. Through bidirectional temporal interaction and contextual semantic enhancement, it models long-term semantic consistency across frames and outputs feature representations with long-term semantic consistency. Multi-scale fusion features are obtained by performing multi-scale processing on long-term semantically consistent feature representations. The progressive multi-scale semantic fusion network module dynamically integrates multi-scale fusion features through adaptive weighting, semantic guidance and low-light enhancement strategies, and outputs multi-scale fusion feature representations. The final change detection result is obtained based on the multi-scale fusion feature representation.

2. The semantic enhancement method for detecting changes in low-light video images according to claim 1, characterized in that, The final change detection result obtained based on the multi-scale fusion feature representation is as follows: The edge-aware multi-scale change detection module detects changes in videos based on multi-scale fusion feature representations, through multi-scale edge detection and an adaptive scale attention mechanism, ensuring target detection performance in low-light environments.

3. The semantic enhancement method for detecting changes in low-light video images according to claim 1, characterized in that, The temporal semantic memory mechanism adopts a bidirectional temporal interaction mechanism to perform forward enhancement calculations. It concatenates and convolves the current time-step features with the next time-step features and then generates weights through the Sigmoid activation function. The forward-enhanced features are obtained by multiplying the original features element-wise; then, the reverse semantic enhancement is performed, using a similar operation process but swapping the order of the features. Finally, the forward-enhanced and reverse-enhanced features are fused through concatenation and convolution operations to achieve the integration of bidirectional temporal information.

4. The semantic enhancement method for detecting changes in low-light video images according to claim 1, characterized in that, The progressive multi-scale semantic fusion network module includes: a semantic guidance module, which generates a semantic attention module with channel-level weights through high-level features, combines a spatial attention module with average pooling and max pooling for target localization, and enhances the performance of low-scale features in low-light environments through a progressive semantic propagation mechanism starting from the deepest features in the network feature pyramid.

5. The semantic enhancement method for detecting changes in low-light video images according to claim 1, characterized in that, The progressive multi-scale semantic fusion network module also includes: The semantic brightness adaptive module adjusts the brightness according to the intensity distribution of semantic features to enhance the semantic visibility of dark areas; The semantic contrast enhancement module improves the discriminability of semantic features through multiplication operations, highlighting the semantic features of small targets; The semantic noise suppression module uses depthwise separable convolution to remove noise while preserving key semantic edge details.

6. The semantic enhancement method for detecting changes in low-light video images according to claim 2, characterized in that, The edge-aware multi-scale change detection module aggregates edge features through edge-aware differential calculation, using Sobel and Laplacian operators, multi-scale convolution kernels, and dilated convolution.

7. A semantic enhancement device for detecting changes in low-light video images, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-6.

Images