Cable tunnel ponding recognition method and device based on improved lightweight deep learning network, terminal equipment and computer readable storage medium

Abstract

The invention discloses a cable tunnel ponding recognition method and device based on an improved lightweight deep learning network, terminal equipment and a computer readable storage medium, and the method comprises the steps: obtaining a to-be-processed cable tunnel ponding recognition image, and inputting the image to a target detection model, the method comprises the following steps of: performing feature extraction on an image by a model through an internal backbone network to obtain a feature information graph, performing feature fusion on the feature information graph through an internal neck network to obtain a plurality of fusion feature graphs, and performing feature enhancement on the selected fusion feature graph through an internal MLCA attention mechanism module to obtain an enhanced fusion feature graph; under the condition that all the fusion feature maps are selected fusion feature maps, performing target detection and classification on the enhanced fusion feature maps through a built-in detection head; otherwise, performing target detection and classification on the enhanced fusion feature map and the fusion feature map which is not enhanced through a built-in detection head, and finally obtaining a recognition object in the image.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a method, device, terminal equipment and computer-readable storage medium for identifying water accumulation in cable tunnels using an improved lightweight deep learning network. Background Art

[0002] Currently, tunnel flooding detection relies primarily on traditional pump and water level detection systems. These systems monitor the water level in a sump and trigger a remote alarm when it exceeds a preset threshold. This indicates the pump's drainage status and guides tunnel maintenance personnel to take emergency measures. However, this traditional method has limitations, as it focuses primarily on the water level in the sump, rather than the overall flooding situation within the tunnel.

[0003] To address this shortcoming, researchers have begun exploring the application of computer vision and deep learning in tunnel water detection. For example, they are using an improved SSD network algorithm to process images captured by cable tunnel inspection robots, employing multi-scale detection methods based on feature maps, and improving the Faster R-CNN method. These methods have achieved some success in tunnel water detection, improving recognition accuracy.

[0004] However, these methods generally employ complex network structures, including multiple layers of convolution, pooling, and full connectivity, as well as numerous parameters and computational complexity. This complex network structure results in large model sizes and places high demands on computing resources and storage space. This is particularly true for resource-constrained embedded robotic devices used in cable tunnel inspections, where limited processors, memory, and storage make it difficult to run large deep learning models in real time. Summary of the Invention

[0005] Embodiments of the present invention provide a method, apparatus, terminal device, and computer-readable storage medium for identifying water accumulation in cable tunnels using an improved lightweight deep learning network, which can reduce the amount of computation and enable the network model to run in real time in a resource-constrained environment.

[0006] An embodiment of the present invention provides a method for identifying water accumulation in a cable tunnel using an improved lightweight deep learning network, comprising:

[0007] Acquire the cable tunnel water accumulation identification image to be processed;

[0008] The cable tunnel water accumulation recognition image to be processed is input into the trained target detection model, so that the target detection model performs feature extraction on the cable tunnel water accumulation recognition image to be processed through the built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through the built-in neck network to obtain several fusion feature maps; the selected fusion feature map is feature enhanced through the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map; when all fusion feature maps are selected fusion feature maps, the enhanced fusion feature map is subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed; otherwise, the enhanced fusion feature map and the unenhanced fusion feature map are subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed;

[0009] The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module, and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module, and the SPPF module are connected in sequence; the second C2f module, the third C2f module, and the SPPF module are connected to the neck network.

[0010] Furthermore, the characteristic information graph includes: a first characteristic information graph, a second characteristic information graph, and a third characteristic information graph;

[0011] Through the built-in backbone network, feature extraction is performed on the cable tunnel water accumulation recognition image to be processed, and a feature information map is obtained, including:

[0012] The first CBS module, the first AKConv module, the first C2f module, the second AKConv module, and the second C2f module in the built-in backbone network are used to extract features from the cable tunnel water accumulation recognition image to be processed, thereby obtaining a first feature information graph.

[0013] Performing feature extraction on the first feature information graph through the third AkConv module and the third C2f module in the built-in backbone network to obtain a second feature information graph;

[0014] The fourth AKConv module, the fourth C2f module and the SPPF module in the built-in backbone network are used to extract features from the second feature information graph to obtain a third feature information graph.

[0015] Furthermore, the plurality of fused feature maps include: a first fused feature map, a second fused feature map, and a third fused feature map;

[0016] A neck network comprising: a first upsampling module, a first splicing module, a fifth C2f module, a second upsampling module, a second splicing module, a sixth C2f module, a second CBS module, a third splicing module, a seventh C2f module, a third CBS module, a fourth splicing module, and an eighth C2f module;

[0017] The feature information graph is fused through the built-in neck network to obtain several fused feature graphs, including:

[0018] After upsampling the third feature information map through the first upsampling module, the third feature information map is spliced with the second feature information map through the first splicing module to obtain a first spliced feature map;

[0019] Perform feature transformation on the first concatenated feature map through the fifth C2f module to obtain a multi-scale feature map;

[0020] After upsampling the multi-scale feature map through the second upsampling module, it is spliced with the first feature information map through the second splicing module to obtain a second spliced feature map;

[0021] Perform feature transformation on the second concatenated feature map through the sixth C2f module to obtain a first fused feature map;

[0022] After convolution and normalization processing are performed on the first fusion feature map by the second CBS module, it is spliced with the multi-scale feature map by the third splicing module to obtain a third spliced feature map;

[0023] Performing feature transformation on the third concatenated feature map through the seventh C2f module to obtain a second fused feature map;

[0024] After convolution and normalization processing is performed on the second fused feature map by the third CBS module, it is spliced with the third feature information map by the fourth splicing module to obtain a fourth spliced feature map;

[0025] The fourth concatenated feature map is transformed by the eighth C2f module to obtain a third fused feature map.

[0026] Furthermore, the selected fusion feature map is enhanced through the built-in MLCA attention mechanism module to obtain the enhanced fusion feature map, including:

[0027] Perform local average pooling on the selected fusion feature map to obtain the local pooled feature map;

[0028] Perform global average pooling on the selected fusion feature map to obtain the globally pooled feature map;

[0029] Performing feature conversion on the local pooled feature map and the global pooled feature map to obtain a converted local feature map and a converted global feature map;

[0030] The converted local feature map and the converted global feature map are subjected to feature fusion and enhancement to obtain an enhanced fused feature map.

[0031] Furthermore, the trained object detection model is determined in the following way:

[0032] Acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and a recognition object in the corresponding image; the recognition object includes any one of the following and a combination thereof: asphalt water accumulation mixture, yellowing scale, and water accumulation;

[0033] Several training data are input into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

[0034] Furthermore, after obtaining some training data, it also includes:

[0035] Performing geometric transformation on the training data to obtain a number of expanded training data; the geometric transformation includes any one of the following and their combination: cropping, rotation, and adding noise.

[0036] Based on the above method embodiment, the present invention provides a corresponding device embodiment, including: an image acquisition module to be identified and a target detection module;

[0037] The module for acquiring an image to be identified is used to acquire an image to be processed for identifying water accumulation in a cable tunnel;

[0038] The target detection module is used to input the cable tunnel water accumulation recognition image to be processed into the trained target detection model, so that the target detection model performs feature extraction on the cable tunnel water accumulation recognition image to be processed through the built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through the built-in neck network to obtain several fusion feature maps; the selected fusion feature map is feature enhanced through the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map; when all fusion feature maps are selected fusion feature maps, the enhanced fusion feature map is subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed; otherwise, the enhanced fusion feature map and the unenhanced fusion feature map are subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed;

[0039] The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module, and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module, and the SPPF module are connected in sequence; the second C2f module, the third C2f module, and the SPPF module are connected to the neck network.

[0040] Furthermore, the target detection module includes: a model training submodule;

[0041] Model training submodule, including: training data acquisition unit and iterative training unit;

[0042] A training data acquisition unit is configured to acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and an identification object in the corresponding image; the identification object includes any one of the following and any combination thereof: asphalt water accumulation mixture, yellowed scale, and water accumulation;

[0043] The iterative training unit is used to input a number of training data into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

[0044] Based on the above-mentioned method embodiment, the present invention provides a corresponding terminal device embodiment, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps of the cable tunnel water accumulation identification method of the improved lightweight deep learning network as described in the present invention.

[0045] Based on the above-mentioned method embodiment, the present invention provides a corresponding computer-readable storage medium embodiment, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to execute the steps of the cable tunnel water accumulation identification method of the improved lightweight deep learning network as described in the present invention.

[0046] Compared with the prior art, the beneficial effects of the embodiment of this solution are:

[0047] The present invention obtains a cable tunnel water accumulation recognition image to be processed, and then inputs the cable tunnel water accumulation recognition image to be processed into a trained target detection model, so that the target detection model performs feature extraction on the cable tunnel water accumulation recognition image to be processed through a built-in backbone network to obtain a feature information graph, wherein the backbone network includes a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module and an SPPF module, and each module is connected in sequence, and the second C2f module, the third C2f module and the SPPF module are connected to the neck network. By introducing the variable kernel convolution AKConv module, the network is allowed to adaptively select convolution kernels of different sizes at different spatial positions, which helps the model better capture the features in the image and may reduce the amount of computation by reducing unnecessary convolution operations; then, the target detection model uses the built-in neck network to fuse the feature information map to obtain several fused feature maps, and uses the built-in MLCA attention mechanism module to enhance the features of the selected fused feature maps to obtain the enhanced fused feature maps. By introducing the MLCA attention mechanism, the model's attention to key features can be enhanced, thereby reducing the computation on irrelevant areas. Finally, when all fused feature maps are selected fused feature maps, the enhanced fused feature maps are used for target detection and classification through the built-in detection head to obtain the identification objects in the cable tunnel water accumulation recognition image to be processed; otherwise, the enhanced fused feature maps and the unenhanced fused feature maps are used for target detection and classification through the built-in detection head to obtain the identification objects in the cable tunnel water accumulation recognition image to be processed. This processing method enables the model to flexibly adjust the complexity of the calculation. When resources are limited or real-time processing is required, only some feature maps can be selected for enhancement to reduce the calculation complexity and increase the processing speed. When there are higher requirements for recognition accuracy and sufficient computing resources, all feature maps can be enhanced to improve the recognition accuracy.

[0048] In summary, the present invention simplifies the network structure through the variable kernel convolution AKConv module in the backbone network and introduces the MLCA attention mechanism to reduce the amount of computation, so that the network model can run in real time in a resource-constrained environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a method for identifying water accumulation in cable tunnels using an improved lightweight deep learning network according to an embodiment of the present invention;

[0050] Figure 2 This is a flow chart of the target detection model provided by one embodiment of the present invention performing target detection and classification tasks;

[0051] Figure 3 1 is a schematic diagram of the structure of an AKConv module provided by an embodiment of the present invention;

[0052] Figure 4 Schematic diagram of the structure of the MLCA attention mechanism module provided by one embodiment of the present invention;

[0053] Figure 5 is a network structure diagram of a target detection model provided by one embodiment of the present invention;

[0054] Figure 6 is a network structure diagram of another target detection model provided by one embodiment of the present invention;

[0055] Figure 7 Schematic diagram of the output result of the target detection model provided by one embodiment of the present invention;

[0056] Figure 8 Schematic diagram of iterative changes in performance indicators during the training of an object detection model provided by one embodiment of the present invention;

[0057] Figure 9 It is a structural schematic diagram of a cable tunnel water accumulation identification device based on an improved lightweight deep learning network provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features.

[0060] like Figure 1 As shown, an embodiment of the present invention provides a method for identifying water accumulation in cable tunnels using an improved lightweight deep learning network, the method comprising at least the following steps:

[0061] Step S1: obtaining a cable tunnel water accumulation recognition image to be processed;

[0062] For step S1, a camera pre-installed in the cable tunnel is used to capture images or videos of accumulated water in the cable tunnel. If the camera records in the form of video, the video needs to be framed to obtain a cable tunnel water accumulation identification image to be processed.

[0063] Step S2: inputting the to-be-processed cable tunnel water accumulation recognition image into the trained target detection model, so that the target detection model performs feature extraction on the to-be-processed cable tunnel water accumulation recognition image through the built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through the built-in neck network to obtain several fusion feature maps; performs feature enhancement on the selected fusion feature map through the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map; when all fusion feature maps are selected fusion feature maps, performs target detection and classification on the enhanced fusion feature map through the built-in detection head to obtain the recognition object in the to-be-processed cable tunnel water accumulation recognition image; otherwise, performs target detection and classification on the enhanced fusion feature map and the unenhanced fusion feature map through the built-in detection head to obtain the recognition object in the to-be-processed cable tunnel water accumulation recognition image;

[0064] The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module, and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module, and the SPPF module are connected in sequence; the second C2f module, the third C2f module, and the SPPF module are connected to the neck network.

[0065] For step S2, the cable tunnel water accumulation recognition image to be processed obtained in step S1 is input into a pre-trained target detection model to perform target detection and classification tasks to obtain the recognition objects in the cable tunnel water accumulation recognition image to be processed. The target detection model has been pre-trained and can recognize the recognition objects in the cable tunnel water accumulation image. Its recognition range covers three main recognition objects: asphalt water accumulation mixture, yellowing scale and water accumulation. It should be noted that the recognition results of the model are not limited to a single type of the above-mentioned recognition objects. In actual applications, the water accumulation situation in the cable tunnel may be quite complicated. Any of the above-mentioned recognition objects may appear alone in the image, or any combination of them may exist at the same time. For example, some areas may only have water accumulation, while other areas may have both asphalt water accumulation mixture and yellowing scale.

[0066] The target detection model uses the YOLOv8 network model as the basic framework, which includes a backbone network (Backbone), a neck network (Neck), an MLCA attention mechanism module and a detection head (Detect). The backbone network includes the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module and the SPPF module; the neck network includes the first upsampling module, the first splicing module, the fifth C2f module, the second upsampling module, the second splicing module, the sixth C2f module, the second CBS module, the third splicing module, the seventh C2f module, the third CBS module, the fourth splicing module and the eighth C2f module; the detection head includes the first detection head, the second detection head and the third detection head.

[0067] Among them, the internal connection relationship between the backbone network and the neck network is as follows: the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module and the SPPF module in the backbone network are connected in sequence; the first upsampling module, the first splicing module, the fifth C2f module, the second upsampling module, the second splicing module and the sixth C2f module in the neck network are connected in sequence, the second CBS module, the third splicing module, the seventh C2f module, the third CBS module, the fourth splicing module and the eighth C2f module are connected in sequence, the fifth C2f module is connected to the third splicing module, and the sixth C2f module is connected to the second CBS module.

[0068] The cross-network connection relationship between the backbone network, neck network, MLCA attention mechanism module and detection head is as follows: the second C2f module in the backbone network is connected to the second splicing module in the neck network, the third C2f module in the backbone network is connected to the first splicing module in the neck network, the SPPF module in the backbone network is connected to the first upsampling module in the neck network, and the SPPF module in the backbone network is connected to the fourth splicing module in the neck network; the eighth C2f module in the neck network is connected to the MLCA attention mechanism module, and the MLCA attention mechanism module is connected to the third detection head; the seventh C2f module in the neck network is connected to the second detection head; the sixth C2f module in the neck network is connected to the first detection head.

[0069] like Figure 2 As shown in the figure, the specific process of the target detection model to perform target detection and classification tasks includes the following steps:

[0070] Step S21: extracting features from the cable tunnel water accumulation recognition image to be processed through the built-in backbone network to obtain a feature information map;

[0071] In a preferred embodiment, the characteristic information graph includes: a first characteristic information graph, a second characteristic information graph, and a third characteristic information graph;

[0072] Through the built-in backbone network, feature extraction is performed on the cable tunnel water accumulation recognition image to be processed, and a feature information map is obtained, including:

[0073] The first CBS module, the first AKConv module, the first C2f module, the second AKConv module, and the second C2f module in the built-in backbone network are used to extract features from the cable tunnel water accumulation recognition image to be processed, thereby obtaining a first feature information graph.

[0074] Performing feature extraction on the first feature information graph through the third AkConv module and the third C2f module in the built-in backbone network to obtain a second feature information graph;

[0075] The fourth AKConv module, the fourth C2f module and the SPPF module in the built-in backbone network are used to extract features from the second feature information graph to obtain a third feature information graph.

[0076] For step S21, first, preliminary feature extraction is performed on the cable tunnel water accumulation recognition image to be processed through the first CBS module, the first AKConv module, the first C2f module, the second AKConv module and the second C2f module in the built-in backbone network. These modules work together on the input image to extract basic feature information and form a first feature information graph. Specifically, the cable tunnel water accumulation recognition image to be processed is subjected to feature extraction through the first CBS module in the backbone network built into the target detection model, wherein the CBS module is a combination module of convolution (Conv), batch normalization (BatchNormalization) and activation function (SiLU), which is used to perform preliminary feature extraction and nonlinear transformation on the image. Subsequently, the image processed by the first CBS module is subjected to further feature extraction through the first AKConv module, wherein, as Figure 3Figure 1 shows the structure of the AKConv module. The AKConv module is an adaptive kernel convolution module that uses a learned offset to adjust the initial sampling shape. The convolution kernel's sampling shape and parameters are adaptively adjusted based on the content of the input feature map and task requirements. The feature map is then resampled based on the adjusted sampling shape. The resampled feature map undergoes reshaping, reconvolution, and normalization, and finally outputs the feature extraction results through an activation function. This adaptive mechanism enables the AKConv module to linearly increase or decrease the number of convolution parameters based on the hardware environment. In simple recognition tasks, the AKConv module reduces the number of convolution parameters to reduce computational complexity and improve processing speed. In complex recognition tasks, the AKConv module increases the number of convolution parameters to capture more feature information to ensure accuracy. The image processed by the first AKConv module then passes through the first C2f module for further feature fusion processing. The C2f module comprises a convolutional layer, a pooling layer, and a feature concatenation layer, which integrates and refines feature information at different levels to generate a more representative feature representation.

[0077] After processing by the first CBS module, the first AKConv module, and the first C2f module, the image already contains a certain level of feature information. However, these features are still not deep or comprehensive enough. Therefore, the image processed by the first C2f module is passed through the second AKConv module to adaptively adjust the sampling shape and parameters of the convolution kernel to further extract deeper feature information, refine and enhance the features, and thus improve the accuracy of target detection. Next, the image processed by the second AKConv module is passed through the second C2f module to further refine and filter the features. By removing redundant and noisy information, key and useful features are retained, generating a more compact and efficient feature representation. The final output is the first feature information map, i.e., the low-level feature map. It mainly contains low-level features in the image, such as edges, texture, and basic shape information. These features are the basis for subsequent higher-level feature extraction.

[0078] Secondly, through the third AkConv module and the third C2f module in the built-in backbone network, deeper feature extraction is performed on the first feature information map. These two modules work together to further refine and enhance the features in the first feature information map to form a second feature information map. Compared with the first feature information map, the second feature information map contains a deeper and more refined feature representation. Specifically, the first feature information map is subjected to feature extraction through the third AkConv module. The module uses adaptive convolution kernel (AKConv) technology to dynamically adjust the parameters of the convolution kernel according to the different input features to more effectively capture the key features in the first feature information map. After the third AkConv module performs a convolution operation on the first feature information map, the output feature map is passed to the third C2f module for feature fusion processing to obtain a second feature information map with a deeper and more refined feature representation, that is, an intermediate feature map, which mainly contains more complex shapes, local structures, and object parts. Information.

[0079] Finally, the fourth AKConv module, the fourth C2f module and the SPPF module in the built-in backbone network are used to further extract and refine the features of the second feature information map. These three modules work together to further refine and enhance the features in the second feature information map to form a third feature information map. Compared with the first feature information map, the second feature information map contains a higher-level and more representative feature representation. Specifically, after the second feature information map is subjected to feature extraction and feature fusion processing by the fourth AKConv module and the fourth C2f module, it is processed by the spatial pyramid pooling structure in the SPPF (Spatial Pyramid Pooling Fast) module, and pooling operations are performed at different scales to capture the multi-scale information in the feature map, further enhancing the representation ability of the feature map. Finally, the third feature information map, i.e., the high-level feature map, is obtained, which mainly contains information such as the global structure, contextual information, and semantic meaning of the object.

[0080] Step S22: performing feature fusion on the feature information graph through the built-in neck network to obtain a plurality of fused feature graphs;

[0081] In a preferred embodiment, the plurality of fused feature maps include: a first fused feature map, a second fused feature map, and a third fused feature map;

[0082] The feature information graph is fused through the built-in neck network to obtain several fused feature graphs, including:

[0083] After upsampling the third feature information map through the first upsampling module, the third feature information map is spliced with the second feature information map through the first splicing module to obtain a first spliced feature map;

[0084] Perform feature transformation on the first concatenated feature map through the fifth C2f module to obtain a multi-scale feature map;

[0085] After upsampling the multi-scale feature map through the second upsampling module, it is spliced with the first feature information map through the second splicing module to obtain a second spliced feature map;

[0086] Perform feature transformation on the second concatenated feature map through the sixth C2f module to obtain a first fused feature map;

[0087] After convolution and normalization processing are performed on the first fusion feature map by the second CBS module, it is spliced with the multi-scale feature map by the third splicing module to obtain a third spliced feature map;

[0088] Performing feature transformation on the third concatenated feature map through the seventh C2f module to obtain a second fused feature map;

[0089] After convolution and normalization processing are performed on the second fused feature map by the third CBS module, the second fused feature map is spliced with the third feature information map by the fourth splicing module to obtain a fourth spliced feature map;

[0090] The fourth concatenated feature map is transformed by the eighth C2f module to obtain a third fused feature map.

[0091] In step S22, the feature information map is subjected to multi-level and multi-scale feature fusion through the built-in neck network, aiming to extract and integrate feature information from different levels and scales to improve the performance of tasks such as target detection or image recognition.

[0092] Specifically, the third feature information map (high-level feature map) is appropriately amplified by the first upsampling module, and the resolution of the third feature information map (high-level feature map) is adjusted so that it is consistent with the resolution of the second feature information map (intermediate feature map). Subsequently, the amplified high-level features and intermediate features are seamlessly spliced through the first splicing module to form a first spliced feature map containing rich information. This feature map not only contains the global information and semantic information of the high-level features, but also integrates the local details and texture information of the intermediate features. Then, after processing by the fifth C2f module, the information in the first spliced feature map is further refined and integrated to generate multi-scale and hierarchical features, namely a multi-scale feature map. This feature map not only contains the information in the original feature map, but also introduces a new feature representation and hierarchical structure through feature transformation. Next, the multi-scale feature map is amplified by the second upsampling module to maintain the same resolution as the first feature information map (low-level feature map) for subsequent feature splicing and fusion. Then, the amplified multi-scale features and the first feature information map (low-level feature map) are seamlessly spliced together by the second splicing module to form a second spliced feature map containing rich information. This feature map not only contains the original information and details of the low-level feature map, but also integrates the global information and semantic information in the multi-scale feature map. Finally, the sixth C2f module performs an in-depth feature transformation on the second spliced feature map to generate the first fused feature map.

[0093] Based on the first fused feature map, the second CBS module extracts features from the first fused feature map and further extracts local structural features from the first fused feature map. These features may include underlying features such as edges, textures, and shapes in the image, as well as higher-level semantic features and contextual information to enhance the overall feature expression capability. Next, the first fused feature map that has passed through the second CBS module is spliced with the multi-scale feature map through the third splicing module to form a new third spliced feature map containing richer information. The multi-scale feature map contains image information of high-level and mid-level features. The splicing not only retains the key information and potential features in the first fused feature map, but also greatly enriches the feature representation of the third spliced feature map by introducing high-level and mid-level features in the multi-scale feature map. Finally, the seventh C2f module extracts and transforms the third spliced feature map to generate the second fused feature map.

[0094] Based on the second fused feature map, the third CBS module performs feature extraction on the second fused feature map, further extracting local structural features from the second fused feature map. These features may include objects, scenes, and their relationships in the image, reflecting the deep meaning of the image content. Next, the second fused feature map that has passed through the third CBS module is spliced with the third feature information map through the fourth splicing module to form a new, more information-rich fourth splicing feature map. By introducing information such as the global structure, contextual information, and semantic meaning of the objects in the third feature information map, the content of the feature representation is greatly enriched. Finally, the eighth C2f module performs feature extraction and transformation on the fourth splicing feature map to generate the third fused feature map.

[0095] Step S23: Perform feature enhancement on the selected fusion feature map through the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map;

[0096] In a preferred embodiment, the selected fusion feature map is enhanced by the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map, including:

[0097] Perform local average pooling on the selected fusion feature map to obtain the local pooled feature map;

[0098] Perform global average pooling on the selected fusion feature map to obtain the globally pooled feature map;

[0099] Performing feature conversion on the local pooled feature map and the global pooled feature map to obtain a converted local feature map and a converted global feature map;

[0100] The converted local feature map and the converted global feature map are subjected to feature fusion and enhancement to obtain an enhanced fused feature map.

[0101] In step S23, in order to improve the performance of the model, the MLCA (Multi-Level Context Attention) attention mechanism module is used to enhance the features of the selected fusion feature map, thereby obtaining a more discriminative feature representation. Specifically, Figure 4The figure shows the structure of the MLCA attention mechanism module. First, a local average pooling (LAP) operation is performed on the selected input fused feature map (CWH). Local average pooling performs an average calculation within a local region of the feature map, capturing local structure and details within the feature map, thereby enhancing the model's understanding of local features. Next, a global average pooling (GAP) operation is performed on the selected fused feature map. Global average pooling performs an average calculation across the entire feature map, capturing global information and context within the feature map. This provides a global perspective for subsequent feature transformation and helps the model better understand the overall image content. Subsequently, the local pooled feature map and the global pooled feature map are transformed using a 1D convolution (Conv1d) to obtain the transformed local feature map and the transformed global feature map. 1D convolution is used to compress feature channels while maintaining the spatial dimensionality. After 1D convolution, the dimensionality of the feature map may change, so the features need to be reshaped to adapt to subsequent operations. After the features are rearranged, the converted local feature maps are combined with the original input features through a "multiplication" operation to strengthen the parts of the original features that are useful for the recognition task. This process is equivalent to a feature selection, allowing the model to focus more on key information and enhance the model's ability to capture useful features. At the same time, the converted global feature maps are combined with local pooling features through an "addition" operation to incorporate global context information into local features, achieving a deep fusion of local and global features. Finally, the feature maps that have undergone local and global attention processing are restored to their original spatial dimensions through an unpooling (UNAP) operation, ensuring the consistency of the enhanced feature maps with the original input feature maps in spatial dimensions, and ultimately obtaining the enhanced fused feature maps.

[0102] The present invention uses the MLCA attention mechanism module to perform local average pooling and global average pooling operations, respectively capturing local structure and detail information in the feature map, as well as global information and contextual relationships. Furthermore, by adopting lightweight pooling operations and 1D convolution, a large amount of computational overhead is avoided, enabling the MLCA module to optimize computational efficiency while maintaining high performance.

[0103] Step S24: When all fused feature maps are selected fused feature maps, the enhanced fused feature maps are subjected to target detection and classification by the built-in detection head to obtain the identification object in the cable tunnel water accumulation recognition image to be processed; otherwise, the enhanced fused feature maps and the unenhanced fused feature maps are subjected to target detection and classification by the built-in detection head to obtain the identification object in the cable tunnel water accumulation recognition image to be processed;

[0104] For step S24, when all fused feature maps are selected fused feature maps, as shown in FIG. Figure 5 As shown in the figure, MLCA attention mechanism modules are set in front of the first, second, and third detection heads to enhance the features of the fused feature map to improve the ability to capture key information. The fused feature map enhanced by the MLCA attention mechanism module is then sent to the corresponding detection head for target detection and classification. Specifically, the first, second, and third detectors receive and process enhanced fused feature maps. These maps contain various hierarchical features extracted by the backbone and neck networks, including low-level features such as edges, textures, and basic shape information; mid-level features such as more complex shapes, local structures, and object components; and high-level features such as the object's global structure, contextual information, and semantic meaning. The detectors perform classification based on these extracted features. For example, when identifying asphalt-water mixtures, edge features can be used to distinguish the boundary between the mixture and the surrounding environment. When identifying ponded water, a distinct edge forms at the interface between the water surface and the ground or other objects. Asphalt-water mixtures may have unique texture features, such as the distribution and arrangement of asphalt particles. Yellow scale exhibits a specific stain or speckled texture. The asphalt-water mixture, yellow scale, and ponded water are distinguished based on their positional relationship within the image. For example, ponded water is located in low-lying areas of the road, while yellow scale appears on surfaces that have been damp for a long time. Finally, the outputs of the first, second, and third detectors are combined to determine the objects in the cable tunnel water recognition image to be processed.

[0105] In the case that not all fused feature maps are selected fused feature maps, in this embodiment, preferably, Figure 6 As shown, an MLCA attention mechanism module is set in front of the third detector head to perform feature enhancement. This means that only the fusion feature map processed by the third detector head will be enhanced by the MLCA module. At the same time, the first detector head and the second detector head will process the fusion feature map that has not been enhanced by the MLCA module. The process of target detection and classification by the detector head is similar. The first detector head, the second detector head and the third detector head respectively receive and process the first fusion feature map, the second fusion feature map and the enhanced fusion feature map, classify the extracted features of each class, and finally combine the output results of the first detector head, the second detector head and the third detector head to obtain the identification object in the cable tunnel water accumulation identification image to be processed. Figure 7 Shown is a structural diagram of the output results of the target detection model of the present invention.

[0106] It should be noted that each additional MLCA attention mechanism module will increase additional computing overhead. Therefore, when choosing whether to set the MLCA module in front of the detection head, it is necessary to weigh the computing resources and performance requirements. If the computing resources are sufficient and the performance requirements are high, you can consider setting the MLCA module in front of multiple detection heads; if the computing resources are limited, you need to carefully select the number of detection heads for setting the MLCA module. In this embodiment, since the third detection head usually corresponds to deep features and contains more high-order semantic information, the introduction of the MLCA attention mechanism module can more effectively optimize feature representation and improve the performance of target detection and classification. In addition, it should be emphasized that in the embodiment of the present invention, setting the MLCA attention mechanism module in front of the third detection head only represents one of the better setting methods. The MLCA attention mechanism module can be set in any one of the first detection head, the second detection head and the third detection head and their combination. In practical applications, the setting of the MLCA attention mechanism module can be flexibly adjusted according to actual conditions to achieve optimal performance and computing efficiency.

[0107] In a preferred embodiment, the trained object detection model is determined by:

[0108] Acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and a recognition object in the corresponding image; the recognition object includes any one of the following and a combination thereof: asphalt water accumulation mixture, yellowing scale, and water accumulation;

[0109] Several training data are input into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

[0110] In a preferred embodiment, after obtaining a number of training data, the method further includes:

[0111] Performing geometric transformation on the training data to obtain a number of expanded training data; the geometric transformation includes any one of the following and their combination: cropping, rotation, and adding noise.

[0112] In one embodiment of the present invention, recognition images of the cable tunnel are collected by a camera, and these images may contain features such as asphalt-water mixture, yellowed scale and accumulated water. Alternatively, a video of the cable tunnel is captured by a surveillance camera, and the video is frame-processed to obtain a series of recognition images. In order to increase the diversity and robustness of the data set, the images in the basic data set are expanded using methods such as cropping, rotation, and noise addition. Data expansion technology helps the model learn more generalized features and improves its recognition performance in different scenarios. Then, the images in the data set containing asphalt-water mixture, yellowed scale and accumulated water are annotated manually or using the labelimg tool to determine the location and size of the accumulated water area and assign corresponding labels to it. The annotated images are divided into a training set, a test set, and a validation set. In this embodiment, after expansion, a data set of 850 pictures including asphalt water mixture, yellowed scale and water accumulation is included. These pictures are divided into training set, validation set and test set in the ratio of 8:1:1. The training set contains 680 pictures for training the target detection model; the validation set contains 85 pictures for adjusting model parameters during training to avoid overfitting; the test set contains 85 pictures for evaluating the performance of the trained model.

[0113] After the target detection model training is completed, the performance of the model is measured using evaluation indicators such as parameter amount, floating-point operation number (GFLOPs), precision rate, recall rate and mean average precision. The parameter amount is used to measure the total number of parameters in the target detection model of the present invention, the floating-point operation number is used to measure the algorithm complexity of the target detection model of the present invention, the precision rate is used to measure the proportion of correct predictions among all targets predicted by the target detection model of the present invention, the recall rate is used to measure the target ratio of all real targets predicted by the target detection model of the present invention when the target detection model of the present invention is predicted, and the mean average precision is used to measure the detection accuracy of the target detection model of the present invention. As shown in Table 1, compared with the various indicators of the YOLOv8 network model, the target detection model of the present invention has decreased in parameter amount and floating-point operation number, and the file size has been reduced from 6.3MB to 6.1MB, while the precision rate and mean average precision have been improved. Overall, the effect of cable tunnel water accumulation recognition is improved while reducing the parameter amount and complexity of the network.

[0114] Table 1 Comparison of model performance indicators

[0115]

[0116] like Figure 8The figure shows the iterative changes of performance indicators during the training of the target detection model, where: Figure 8 (a) train / box_loss refers to the bounding box loss during training, which is used to measure the difference between the bounding box predicted by the model and the true bounding box; Figure 8 (b) train / cls_loss refers to the classification loss during training, which is used to measure the accuracy of model classification; Figure 8 (c) train / dfl_loss refers to the loss function during training; Figure 8 (g) metrics / precision (B) refers to the precision rate in the evaluation indicators; Figure 8 (h) metrics / recall (B) refers to the recall rate (Recall) in the evaluation indicators; Figure 8 (d) to Figure 8 (f) val / box_loss, val / cls_loss, and val / dfl_loss correspond to the corresponding indicators during training and are used for the validation set; Figure 8 (i) metrics / mAP50(B) and Figure 8 (j)metrics / mAP50-95(B) refers to the mean average precision (MAP) in the evaluation indicators. mAP50 represents the average precision when the intersection over union (IoU) threshold is 0.5, while mAP50-95 represents the average of the average precisions when the IoU threshold is between 0.5 and 0.95.

[0117] like Figure 9 As shown, based on the above method embodiment, a corresponding device embodiment is provided;

[0118] An embodiment of the present invention provides a device for identifying water accumulation in a cable tunnel using an improved lightweight deep learning network, comprising: an image acquisition module to be identified and a target detection module;

[0119] The module for acquiring an image to be identified is used to acquire an image to be processed for identifying water accumulation in a cable tunnel;

[0120] The target detection module is used to input the cable tunnel water accumulation recognition image to be processed into the trained target detection model, so that the target detection model performs feature extraction on the cable tunnel water accumulation recognition image to be processed through the built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through the built-in neck network to obtain several fusion feature maps; the selected fusion feature map is feature enhanced through the built-in MLCA attention mechanism module to obtain an enhanced fusion feature map; when all fusion feature maps are selected fusion feature maps, the enhanced fusion feature map is subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed; otherwise, the enhanced fusion feature map and the unenhanced fusion feature map are subjected to target detection and classification through the built-in detection head to obtain the recognition object in the cable tunnel water accumulation recognition image to be processed;

[0121] The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module, and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module, and the SPPF module are connected in sequence; the second C2f module, the third C2f module, and the SPPF module are connected to the neck network.

[0122] In a preferred embodiment, the target detection module includes: a model training submodule;

[0123] Model training submodule, including: training data acquisition unit and iterative training unit;

[0124] A training data acquisition unit is configured to acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and an identification object in the corresponding image; the identification object includes any one of the following and any combination thereof: asphalt water accumulation mixture, yellowed scale, and water accumulation;

[0125] The iterative training unit is used to input a number of training data into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

[0126] It can be understood that the above-mentioned device embodiment corresponds to the method embodiment of the present invention, which can implement the cable tunnel water accumulation identification method of the improved lightweight deep learning network provided by any one of the above-mentioned method embodiments of the present invention.

[0127] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. Furthermore, in the drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which may be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement the present invention without inventive effort.

[0128] Based on the above-mentioned embodiment of the improved cable tunnel water accumulation identification method using a lightweight deep learning network, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the improved cable tunnel water accumulation identification method using a lightweight deep learning network according to any embodiment of the present invention.

[0129] For example, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more module elements may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.

[0130] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0131] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, connecting various parts of the entire terminal device using various interfaces and lines.

[0132] Based on the above method embodiment, another embodiment is provided: another embodiment of the present invention provides a computer-readable storage medium, including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the cable tunnel water accumulation identification method of the improved lightweight deep learning network described in any one of the above method embodiments of the present invention.

[0133] Among them, the module / unit integrated into the cable tunnel water accumulation identification device / terminal equipment of the improved lightweight deep learning network, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0134] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying water accumulation in cable tunnels based on an improved lightweight deep learning network, characterized in that: include: Acquire the cable tunnel water accumulation identification image to be processed; The cable tunnel water accumulation recognition image to be processed is input into a trained target detection model, so that the target detection model performs feature extraction on the cable tunnel water accumulation recognition image to be processed through a built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through a built-in neck network to obtain a plurality of fused feature maps; the selected fused feature map is feature enhanced through a built-in MLCA attention mechanism module to obtain an enhanced fused feature map; when all fused feature maps are selected fused feature maps, the enhanced fused feature map is subjected to target detection and classification through a built-in detection head to obtain an identification object in the cable tunnel water accumulation recognition image to be processed; Otherwise, the enhanced fusion feature map and the unenhanced fusion feature map are used for target detection and classification through the built-in detection head to obtain the identification object in the cable tunnel water accumulation recognition image to be processed; The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module and the SPPF module are connected in sequence; the second C2f module, the third C2f module and the SPPF module are connected to the neck network.

2. The cable tunnel water accumulation identification method based on an improved lightweight deep learning network according to claim 1 is characterized in that: The characteristic information graph includes: a first characteristic information graph, a second characteristic information graph, and a third characteristic information graph; The said backbone network is used to extract features from the cable tunnel water accumulation recognition image to be processed to obtain a feature information map, including: The first CBS module, the first AKConv module, the first C2f module, the second AKConv module, and the second C2f module in the built-in backbone network are used to extract features of the cable tunnel water accumulation recognition image to be processed, thereby obtaining a first feature information graph; Performing feature extraction on the first feature information graph through a third AkConv module and a third C2f module in the built-in backbone network to obtain a second feature information graph; The fourth AKConv module, the fourth C2f module and the SPPF module in the built-in backbone network are used to extract features from the second feature information graph to obtain a third feature information graph.

3. The cable tunnel water accumulation identification method based on the improved lightweight deep learning network according to claim 2 is characterized in that: The plurality of fused feature maps include: a first fused feature map, a second fused feature map, and a third fused feature map; The neck network includes: a first upsampling module, a first splicing module, a fifth C2f module, a second upsampling module, a second splicing module, a sixth C2f module, a second CBS module, a third splicing module, a seventh C2f module, a third CBS module, a fourth splicing module and an eighth C2f module; The feature information graph is subjected to feature fusion by the built-in neck network to obtain several fused feature graphs, including: After upsampling the third feature information map by the first upsampling module, the third feature information map is spliced with the second feature information map by the first splicing module to obtain a first spliced feature map; Performing feature transformation on the first spliced feature map through the fifth C2f module to obtain a multi-scale feature map; After upsampling the multi-scale feature map through the second upsampling module, the multi-scale feature map is spliced with the first feature information map through the second splicing module to obtain a second spliced feature map; Performing feature transformation on the second spliced feature map through the sixth C2f module to obtain the first fused feature map; After performing convolution and normalization processing on the first fused feature map through the second CBS module, the first fused feature map is spliced with the multi-scale feature map through the third splicing module to obtain a third spliced feature map; Performing feature transformation on the third spliced feature map through the seventh C2f module to obtain the second fused feature map; After performing convolution and normalization processing on the second fused feature map through the third CBS module, the second fused feature map is spliced with the third feature information map through the fourth splicing module to obtain a fourth spliced feature map; The eighth C2f module performs feature transformation on the fourth spliced feature map to obtain the third fused feature map.

4. The cable tunnel water accumulation identification method based on an improved lightweight deep learning network according to claim 3 is characterized in that: The selected fusion feature map is enhanced through the built-in MLCA attention mechanism module to obtain the enhanced fusion feature map, including: Perform local average pooling on the selected fusion feature map to obtain the local pooled feature map; Perform global average pooling on the selected fusion feature map to obtain the globally pooled feature map; Performing feature conversion on the local pooled feature map and the global pooled feature map to obtain a converted local feature map and a converted global feature map; The converted local feature map and the converted global feature map are subjected to feature fusion and enhancement to obtain an enhanced fused feature map.

5. The cable tunnel water accumulation identification method based on an improved lightweight deep learning network according to claim 1 is characterized in that: The trained target detection model is determined by: Acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and a recognition object in the corresponding image; the recognition object includes any one of the following and a combination thereof: asphalt water accumulation mixture, yellowing scale, and water accumulation; Several training data are input into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

6. The cable tunnel water accumulation identification method based on an improved lightweight deep learning network according to claim 5 is characterized in that: After obtaining some training data, it also includes: Performing geometric transformation on the training data to obtain a number of expanded training data; the geometric transformation includes any one of the following and their combination: cropping, rotation, and adding noise.

7. A cable tunnel water accumulation identification device based on an improved lightweight deep learning network, characterized in that: include: Image acquisition module to be identified and target detection module; The to-be-identified image acquisition module is used to acquire a to-be-processed cable tunnel water accumulation recognition image; The target detection module is used to input the to-be-processed cable tunnel water accumulation recognition image into a trained target detection model, so that the target detection model performs feature extraction on the to-be-processed cable tunnel water accumulation recognition image through a built-in backbone network to obtain a feature information map, and performs feature fusion on the feature information map through a built-in neck network to obtain a plurality of fused feature maps; performs feature enhancement on the selected fused feature map through a built-in MLCA attention mechanism module to obtain an enhanced fused feature map; when all fused feature maps are selected fused feature maps, performs target detection and classification on the enhanced fused feature map through a built-in detection head to obtain an identification object in the to-be-processed cable tunnel water accumulation recognition image; Otherwise, the enhanced fusion feature map and the unenhanced fusion feature map are used for target detection and classification through the built-in detection head to obtain the identification object in the cable tunnel water accumulation recognition image to be processed; The backbone network includes: a first CBS module, a first AKConv module, a first C2f module, a second AKConv module, a second C2f module, a third AkConv module, a third C2f module, a fourth AKConv module, a fourth C2f module and an SPPF module; wherein the first CBS module, the first AKConv module, the first C2f module, the second AKConv module, the second C2f module, the third AkConv module, the third C2f module, the fourth AKConv module, the fourth C2f module and the SPPF module are connected in sequence; the second C2f module, the third C2f module and the SPPF module are connected to the neck network.

8. The cable tunnel water accumulation identification device based on an improved lightweight deep learning network according to claim 7 is characterized in that: The target detection module includes: a model training submodule; The model training submodule includes: a training data acquisition unit and an iterative training unit; The training data acquisition unit is used to acquire a plurality of training data; each training data includes: a cable tunnel water accumulation recognition image and an identification object in the corresponding image; the identification object includes any one of the following and a combination thereof: asphalt water accumulation mixture, yellowing scale, and water accumulation; The iterative training unit is used to input a number of training data into the target detection model to be trained for iterative training until the loss function converges to obtain a trained target detection model.

9. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the method for identifying water accumulation in a cable tunnel using an improved lightweight deep learning network as described in any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium, characterized in that include: A stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the cable tunnel water accumulation identification method based on the improved lightweight deep learning network as described in any one of claims 1 to 6.

Images