Power scene water leakage and oil leakage monitoring method based on instance segmentation, medium and equipment

Abstract

The invention provides a power scene water leakage and oil leakage monitoring method based on instance segmentation, a medium and equipment. The method is realized through a scene monitoring model. The scene monitoring model comprises a backbone network, a neck structure and a head structure; in the backbone network, an input image is processed by a CBS convolution block sequence, is processed by a plurality of stages, and is processed by a spatial pyramid pooling fast module SPPF and a cross-level pyramid slice attention module C2PSA; each stage comprises CBS convolution block sequence processing and processing executed by a wavelet convolution converter; and the neck structure performs feature fusion and splicing on the outputs of the plurality of stages and the final output of the backbone network, and then inputs into the head structure to obtain a water leakage and oil leakage monitoring result. According to the method, local-global and high-frequency-low-frequency feature interaction of information of the water leakage and oil leakage region can be enhanced, and accurate segmentation of the water leakage and oil leakage region is realized, so that the monitoring accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more specifically to a method, medium and device for monitoring water and oil leakage in power scenes based on instance segmentation. Background Art

[0002] In the field of power generation, detecting water and oil leaks in power plant equipment is a crucial step in ensuring safe operation. Existing technologies primarily rely on manual inspections, sensor networks, and traditional image processing methods. However, these technologies have significant shortcomings in detection accuracy, real-time performance, and environmental adaptability. A detailed analysis is provided below: (1) Water and oil leak identification methods based on manual inspections and sensor networks: Traditional detection methods are mainly manual inspections, supplemented by sensors arranged in point or short line segments (such as radar level gauges, infrared cameras, etc.); for example, hydropower plant buildings often monitor water leaks through manual observation and local sensors, but the coverage is limited, hidden areas are easily missed, and the response delay is significant. Some scenarios use sensor networks such as oil leakage ropes and flow monitoring equipment, but they need to be densely deployed around leak-prone points (such as oil tanks and pipelines), which is costly and complex to maintain, and the detection ability of small leaks or complex pipeline structures is insufficient; (2) Visual methods based on traditional image processing: In recent years, video surveillance and image recognition technologies have been introduced to power plant water and oil leak detection. For example, infrared cameras are used to capture areas with abnormal temperatures, or visible light cameras are combined with background subtraction algorithms such as the Bernsen threshold method to extract candidate leak areas. However, these methods have poor robustness under complex lighting conditions and interference from liquid reflections, and rely on manually designed features, making it difficult to accurately locate the target area. (3) Improved Methods Based on Deep Learning: Some studies have attempted to improve detection accuracy by incorporating deep learning. For example, the EfficientDet network was used to classify leaking liquids from pipelines, alleviating data shortages through automatic data augmentation. However, the imbalance between normal and leaking samples in the training data resulted in low recognition rates for the minority class. Furthermore, existing methods often employ object detection frameworks (such as YOLO and Faster R-CNN), which can only select leaking areas without pixel-level segmentation, making precise localization difficult.

[0003] The C3k2 structure used in the yolov11-seg backbone network is a typical residual convolution module, which extracts features through stacked convolutions and relies on local receptive fields. Due to the prevalence of complex interference such as specular reflections and weak texture boundaries in leaking and oil spills, the C3k2 module's feature transfer mechanism based on residual connections struggles to establish effective local-global associations, resulting in the loss of high-resolution detail features. Repeated convolution operations lead to feature redundancy, making it difficult to capture subtle changes in water surface texture. Furthermore, C3k2 has weak global perception capabilities. Its convolution kernels focus only on local neighborhood information and are unable to model long-range dependencies. Large leaks and oil spills require global contextual understanding, while small leaks require precise local positioning. However, traditional convolutions struggle to address multiple scales. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies in the prior art, the purpose of the present invention is to provide a method, medium and equipment for monitoring water and oil leakage in power scenarios based on instance segmentation; this method can enhance the local-global and high-frequency-low-frequency feature interaction of water and oil leakage area information, achieve accurate water and oil leakage area segmentation, and thus improve monitoring accuracy.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for monitoring water and oil leakage in power scenes based on instance segmentation, which is implemented by reasoning on monitoring images collected in the power scene through a scene monitoring model; the scene monitoring model includes: a backbone network, a neck structure, and a head structure; In the backbone network, the input image is processed by a sequence of CBS convolutional blocks, then through several stages of processing, and then processed by a spatial pyramid pooling fast module SPPF and a cross-level pyramid slice attention module C2PSA; Each stage includes CBS convolutional block sequence processing and processing performed by a wavelet convolution transformer; wherein the CBS convolutional block sequence processing refers to: performing convolution, batch normalization, and activation in sequence; the wavelet convolution transformer includes: a first channel consisting of a convolution layer and a two-layer routing feature fusion transformer; and a second channel consisting of a wavelet feature extraction module to enhance the feature interaction between local and global, high-frequency and low-frequency features; the first channel and the second channel are spliced to obtain the output of the wavelet convolution transformer; The neck structure performs feature fusion and splicing on the outputs of multiple stages and the final output of the backbone network, and then inputs them into the head structure; based on the output results of the head structure, the water and oil leakage monitoring results are obtained.

[0006] Preferably, in the dual-layer routing feature fusion transformer, the input of the dual-layer routing feature fusion transformer is sequentially processed by image block embedding and depthwise separable convolution, and the output of the image block embedding is added to the output of the depthwise separable convolution; then, after layer normalization layer 1 and dual-layer routing attention processing, the obtained features are added to the features before layer normalization layer 1 processing; then, after layer normalization layer 2 and multi-layer perceptron processing, the obtained features are added to the features before layer normalization layer 2 processing to obtain the output of the dual-layer routing feature fusion transformer.

[0007] Preferably, in the wavelet feature extraction module, the input of the wavelet feature extraction module is sequentially processed by a CBS convolution block sequence, split, and processed by m wavelet convolution bottleneck layers, the obtained features are concatenated with the split output, and then processed by a CBS convolution block sequence to obtain the output of the wavelet feature extraction module; In the wavelet convolution bottleneck layer, the input of the wavelet convolution bottleneck layer is sequentially subjected to convolution and wavelet convolution, and then added to the input of the wavelet convolution bottleneck layer to obtain the output of the wavelet convolution bottleneck layer.

[0008] Preferably, in the spatial pyramid pooling fast module SPPF, the input features of the spatial pyramid pooling fast module SPPF are sequentially processed by a CBS convolution block sequence and several global maximum pooling layers, the output of the CBS convolution block sequence processing and the output of each global maximum pooling layer are spliced, and then the output of the spatial pyramid pooling fast module SPPF is obtained after being processed by the CBS convolution block sequence.

[0009] Preferably, in the cross-level pyramid slice attention module C2PSA, the input features of the cross-level pyramid slice attention module C2PSA are sequentially processed by CBS convolution block sequence, split and PSABlock block processing, the output of the CBS convolution block sequence processing is spliced with the output of the PSABlock block, and then processed by CBS convolution block sequence to obtain the output of the cross-level pyramid slice attention module C2PSA.

[0010] Preferably, the neck structure includes a cross-scale heterogeneous feature fusion module; the cross-scale heterogeneous feature fusion module is used to fuse the intermediate feature maps of three different sizes: large, medium and small; In the cross-scale heterogeneous feature fusion module, the large-size intermediate feature map is downsampled through global average pooling and global maximum pooling, and the large-size part is obtained through convolution; the medium-size intermediate feature map is convolved to obtain the medium-size part; the small-size intermediate feature map is upsampled and convolved to obtain the small-size part; the large, medium and small-size parts are spliced together to obtain the output of the cross-scale heterogeneous feature fusion module.

[0011] Preferably, there are two cross-scale heterogeneous feature fusion modules, namely: cross-scale heterogeneous feature fusion module 1 and cross-scale heterogeneous feature fusion module 2; The backbone network includes four stages, namely the first stage, the second stage, the third stage and the fourth stage; The three intermediate feature maps of different sizes (large, medium, and small) input to the cross-scale heterogeneous feature fusion module 1 are: the feature map after convolution of the output of the second stage, the feature map after convolution of the output of the third stage, and the feature map after convolution of the final output of the backbone network; The large, medium, and small intermediate feature maps of different sizes input to the cross-scale heterogeneous feature fusion module 2 are: the feature map after convolution of the output of the first stage, the feature map of the output of the second stage, and the feature map obtained by C3k2 feature extraction and convolution of the output of the cross-scale heterogeneous feature fusion module 1; The head structure includes split head 1, split head 2 and split head 3; The output of the cross-scale heterogeneous feature fusion module 2 is extracted by C3k2 features to obtain a large-scale feature map, which is then output to the segmentation head 1; After the large-scale feature map is processed by the CBS convolution block sequence, it is concatenated with the feature map obtained by C3k2 feature extraction of the output of the cross-scale heterogeneous feature fusion module 1, and then subjected to C3k2 feature extraction again to obtain a medium-sized feature map, which is output to the segmentation head 2; The medium-sized feature map is processed by the CBS convolution block sequence and then concatenated with the final output of the backbone network. It is then extracted by C3k2 features to obtain a small-sized feature map, which is output to the segmentation head three.

[0012] Preferably, the C3k2 feature extraction means that the input of the C3k2 feature extraction is sequentially processed by a CBS convolution block sequence, split and processed by n C3k feature extractions, the obtained features are spliced with the split output, and then processed by a CBS convolution block sequence to obtain the output of the C3k2 feature extraction.

[0013] A readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the method for monitoring water and oil leakage in power scenarios based on instance segmentation.

[0014] A computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the method for monitoring water and oil leakage in power scenarios based on instance segmentation is implemented.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. To address the fundamental flaws of traditional CNN models (such as U-Net and DeepLabv3+) in global feature modeling, this paper proposes architectural innovations. The designed wavelet convolution transformer uses a dual-channel design, combining features extracted from wavelet convolutional neural networks and dual-layer routing attention, to enhance the local-global and high-frequency-low-frequency feature interactions of information in leaking and oil spill areas. 2. The traditional FPN fusion mechanism only upsamples small-scale feature maps and then adds them to the previous feature layer, ignoring the rich texture information of the larger feature layer. Repeated convolution operations also over-smoothe high-frequency details (such as water surface edges), resulting in edge information attenuation. To this end, the cross-scale heterogeneous feature fusion module designed in the neck structure of the present invention replaces the upsampling structure in the traditional FPN. By referencing and comparing the shape or appearance changes of water and oil at different scales through multi-scale feature maps, the cross-scale feature fusion module learns the characteristics of the edge segmentation of leaking areas, avoiding repeated convolution operations that over-smoothe high-frequency details (such as fluid edges) and lead to edge information attenuation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a principle block diagram of the scene monitoring model of the present invention; Figure 2 It is a principle block diagram of CBS convolution block sequence processing of the present invention; Figure 3 It is a principle block diagram of the wavelet convolution transformer of the present invention; Figure 4 It is a principle block diagram of the spatial pyramid pooling fast module SPPF of the present invention; Figure 5 This is a block diagram of the principle of the cross-level pyramid slice attention module C2PSA of the present invention; Figure 6 This is a principle block diagram of the cross-scale heterogeneous feature fusion module of the present invention; Figure 7 It is a principle block diagram of the C3k2 feature extraction of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 This embodiment is a method for monitoring water and oil leakage in power scenes based on instance segmentation. It is implemented by reasoning the monitoring images collected in the power scene through the scene monitoring model. The present invention transforms the YOLOv11-seg network architecture, designs a wavelet convolution transformer and a cross-scale heterogeneous feature fusion module, and realizes accurate water and oil leakage area segmentation. Specifically, the scene monitoring model includes: a backbone network, a neck structure and a head structure, such as Figure 1 shown.

[0019] In the backbone network, the input image is processed by a sequence of CBS convolutional blocks, then by several consecutively stacked stages, and then by a spatial pyramid pooling fast module SPPF and a cross-level pyramid slicing attention module C2PSA.

[0020] CBS convolution block sequence processing is used to reduce image resolution to reduce the amount of computation. CBS convolution block sequence processing means: performing convolution, batch normalization and activation in sequence, such as Figure 2 shown.

[0021] Each stage consists of a sequence of CBS convolution blocks and processing performed by a wavelet convolution transformer. The wavelet convolution transformer introduces a wavelet convolution module into Bottleneck. Wavelet convolution uses wavelet transforms to expand the convolution receptive field through multi-frequency response and performs small kernel convolution operations in different frequency ranges. Through wavelet decomposition, the model can capture low-frequency information over a wider range.

[0022] The present invention adopts a wavelet convolution transformer to replace the C3k2 module in the backbone network of yolov11-seg. The wavelet convolution transformer combines the feature extraction capabilities of wavelet convolutional neural network and double-layer routing attention through a dual-channel design to enhance local-global and high-frequency-low-frequency feature interactions.

[0023] The wavelet convolution transformer includes: a first channel consisting of a convolution layer and a two-layer routing feature fusion transformer, and a second channel consisting of a wavelet feature extraction module to enhance the feature interaction between local and global, high frequency and low frequency; the output of the wavelet convolution transformer is obtained by splicing the first channel and the second channel, such as Figure 3 The feature representations of the two channels form a complementary spatial-frequency domain, ultimately achieving effective fusion of multi-scale features.

[0024] In the dual-layer routing feature fusion transformer, the input is processed sequentially through image patch embedding and depthwise separable convolution. The output of the image patch embedding is added to the output of the depthwise separable convolution. The features obtained are then processed through layer-normalization layer 1 and dual-layer routing attention, and the resulting features are added to the features before layer-normalization layer 1. Finally, the features obtained are processed through layer-normalization layer 2 and a multi-layer perceptron, and the resulting features are added to the features before layer-normalization layer 2 to obtain the output of the dual-layer routing feature fusion transformer. The dual-layer routing feature fusion transformer implements dynamic sparse attention to focus on key areas. The dual-layer routing attention adaptively selects the top-K important areas. Through dual-branch computation, coarse-grained routing quickly selects important areas, while fine-grained attention only computes interactions in key areas, reducing computational complexity.

[0025] In the wavelet feature extraction module, the input of the wavelet feature extraction module is processed by the CBS convolution block sequence, split and processed by m wavelet convolution bottleneck layers in sequence. The obtained features are concatenated with the split output and then processed by the CBS convolution block sequence to obtain the output of the wavelet feature extraction module. In the wavelet convolution bottleneck layer, the input of the wavelet convolution bottleneck layer is sequentially subjected to convolution and wavelet convolution, and then added to the input of the wavelet convolution bottleneck layer to obtain the output of the wavelet convolution bottleneck layer.

[0026] The spatial pyramid pooling fast module SPPF is mainly used to fuse multi-scale features. Figure 4 As shown in the figure, in the spatial pyramid pooling fast module SPPF, the input features of the spatial pyramid pooling fast module SPPF are processed in sequence by CBS convolution block sequence and several global maximum pooling layers, the output of the CBS convolution block sequence processing and the output of each global maximum pooling layer are spliced, and then processed by CBS convolution block sequence to obtain the output of the spatial pyramid pooling fast module SPPF.

[0027] The cross-level pyramid slice attention module C2PSA uses multi-scale convolution and channel weighting to improve the model's perception of target details and output the final feature map. Figure 5 As shown, in the cross-level pyramid slice attention module C2PSA, the input features of the cross-level pyramid slice attention module C2PSA are sequentially processed by CBS convolution block sequence, split and PSABlock block processing, the output of the CBS convolution block sequence processing is spliced with the output of the PSABlock block, and then processed by CBS convolution block sequence to obtain the output of the cross-level pyramid slice attention module C2PSA.

[0028] The neck structure fuses and concatenates the outputs of multiple stages and the final output of the backbone network, and then inputs them into the head structure.

[0029] Because different feature layers in the backbone network have different sizes, the traditional FPN fusion mechanism only upsamples the small-sized feature maps and then adds them to the previous layer's features, ignoring the rich texture information in the larger-sized feature layers. Furthermore, repeated 3×3 convolution operations over-smoothe high-frequency details, resulting in attenuation of edge information.

[0030] Therefore, the present invention designs a cross-scale heterogeneous feature fusion module within the neck structure, replacing the upsampling structure used in traditional FPNs. Before feature fusion, this module first adjusts the number of feature channels to align with the primary scale characteristics. The large-scale intermediate feature maps are then downsampled using a hybrid architecture (max pooling + average pooling), helping to preserve the effectiveness and diversity of high-resolution features and leaky and oily images. For the small-scale intermediate feature maps, a convolutional module also adjusts the number of channels, followed by upsampling using the nearest neighbor interpolation method. This helps preserve the rich local features of the low-resolution image and prevents loss of target feature information.

[0031] The neck structure includes a cross-scale heterogeneous feature fusion module. Figure 6 As shown in the figure, the cross-scale heterogeneous feature fusion module is used to fuse the intermediate feature maps of three different sizes: large, medium and small. In the cross-scale heterogeneous feature fusion module, the large-size intermediate feature map is down-sampled by global average pooling and global maximum pooling, and the large-size part is obtained by convolution. The medium-size intermediate feature map is convolved to obtain the medium-size part. The small-size intermediate feature map is subjected to nearest neighbor upsampling and convolution to obtain the small-size part. The large, medium and small size parts are spliced to obtain the output of the cross-scale heterogeneous feature fusion module.

[0032] Specifically, in this embodiment, there are two cross-scale heterogeneous feature fusion modules, namely, cross-scale heterogeneous feature fusion module 1 and cross-scale heterogeneous feature fusion module 2. The backbone network includes four stages, namely, the first stage, the second stage, the third stage, and the fourth stage.

[0033] The three intermediate feature maps of different sizes (large, medium, and small) input to the cross-scale heterogeneous feature fusion module 1 are: the feature map after convolution of the output of the second stage, the feature map after convolution of the output of the third stage, and the feature map after convolution of the final output of the backbone network; The large, medium, and small intermediate feature maps of different sizes input to the cross-scale heterogeneous feature fusion module 2 are: the feature map after convolution of the output of the first stage, the feature map of the output of the second stage, and the feature map obtained by C3k2 feature extraction and convolution of the output of the cross-scale heterogeneous feature fusion module 1; The head structure includes segmentation head one, segmentation head two and segmentation head three; the output of the cross-scale heterogeneous feature fusion module two is extracted by C3k2 features to obtain a large-size feature map, which is output to the segmentation head one; the large-size feature map is processed by the CBS convolution block sequence, and then spliced with the feature map obtained by the output of the cross-scale heterogeneous feature fusion module one after C3k2 feature extraction, and then subjected to C3k2 feature extraction to obtain a medium-size feature map, which is output to the segmentation head two; the medium-size feature map is processed by the CBS convolution block sequence, and then spliced with the final output of the backbone network, and then subjected to C3k2 feature extraction to obtain a small-size feature map, which is output to the segmentation head three.

[0034] like Figure 7 As shown in FIG, C3k2 feature extraction means that the input of C3k2 feature extraction is sequentially processed by a CBS convolution block sequence, split, and processed by n C3k feature extractions, the obtained features are concatenated with the split output, and then processed by a CBS convolution block sequence to obtain the output of C3k2 feature extraction.

[0035] According to the output results of the head structure, the water and oil leakage monitoring results are obtained.

[0036] Example 2 This embodiment provides a readable storage medium, wherein the readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the method for monitoring water and oil leakage in power scenarios based on instance segmentation as described in the first embodiment.

[0037] Example 3 This embodiment provides a computer device, including a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the method for monitoring water and oil leakage in power scenarios based on instance segmentation as described in the first embodiment is implemented.

[0038] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for monitoring water and oil leakage in power scenarios based on instance segmentation, characterized by: This is achieved by reasoning on the monitoring images collected in the power scene through the scene monitoring model; The scene monitoring model includes: a backbone network, a neck structure and a head structure; In the backbone network, the input image is processed by a sequence of CBS convolutional blocks, then through several stages of processing, and then processed by a spatial pyramid pooling fast module SPPF and a cross-level pyramid slice attention module C2PSA; Each stage includes CBS convolutional block sequence processing and processing performed by a wavelet convolution transformer; wherein the CBS convolutional block sequence processing refers to: performing convolution, batch normalization, and activation in sequence; the wavelet convolution transformer includes: a first channel consisting of a convolution layer and a two-layer routing feature fusion transformer; and a second channel consisting of a wavelet feature extraction module to enhance the feature interaction between local and global, high-frequency and low-frequency features; the first channel and the second channel are spliced to obtain the output of the wavelet convolution transformer; The neck structure performs feature fusion and splicing on the outputs of multiple stages and the final output of the backbone network, and then inputs them into the head structure; based on the output results of the head structure, the water and oil leakage monitoring results are obtained.

2. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 1 is characterized by: In the dual-layer routing feature fusion transformer, the input of the dual-layer routing feature fusion transformer is sequentially processed by image block embedding and depthwise separable convolution, and the output of the image block embedding is added to the output of the depthwise separable convolution; then, the input is processed by layer normalization layer 1 and dual-layer routing attention, and the obtained features are added to the features before layer normalization layer 1 processing; then, the input is processed by layer normalization layer 2 and multi-layer perceptron, and the obtained features are added to the features before layer normalization layer 2 processing to obtain the output of the dual-layer routing feature fusion transformer.

3. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 1 is characterized by: In the wavelet feature extraction module, the input of the wavelet feature extraction module is sequentially processed by a CBS convolution block sequence, split and processed by m wavelet convolution bottleneck layers, the obtained features are concatenated with the split output, and then processed by a CBS convolution block sequence to obtain the output of the wavelet feature extraction module; In the wavelet convolution bottleneck layer, the input of the wavelet convolution bottleneck layer is sequentially subjected to convolution and wavelet convolution, and then added to the input of the wavelet convolution bottleneck layer to obtain the output of the wavelet convolution bottleneck layer.

4. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 1 is characterized by: In the spatial pyramid pooling fast module SPPF, the input features of the spatial pyramid pooling fast module SPPF are sequentially processed by a CBS convolution block sequence and several global maximum pooling layers, the output of the CBS convolution block sequence processing and the output of each global maximum pooling layer are spliced, and then the output of the spatial pyramid pooling fast module SPPF is obtained after being processed by the CBS convolution block sequence.

5. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 1 is characterized by: In the cross-level pyramid slice attention module C2PSA, the input features of the cross-level pyramid slice attention module C2PSA are sequentially processed by CBS convolution block sequence, split and PSABlock block, the output of the CBS convolution block sequence processing is spliced with the output of the PSABlock block, and then processed by CBS convolution block sequence to obtain the output of the cross-level pyramid slice attention module C2PSA.

6. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 1 is characterized by: The neck structure includes a cross-scale heterogeneous feature fusion module; the cross-scale heterogeneous feature fusion module is used to fuse the intermediate feature maps of three different sizes: large, medium and small; In the cross-scale heterogeneous feature fusion module, the large-size intermediate feature map is downsampled through global average pooling and global maximum pooling, and the large-size part is obtained through convolution; the medium-size intermediate feature map is convolved to obtain the medium-size part; the small-size intermediate feature map is upsampled and convolved to obtain the small-size part; the large, medium and small-size parts are spliced together to obtain the output of the cross-scale heterogeneous feature fusion module.

7. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 6 is characterized by: There are two cross-scale heterogeneous feature fusion modules, namely: cross-scale heterogeneous feature fusion module 1 and cross-scale heterogeneous feature fusion module 2; The backbone network includes four stages, namely the first stage, the second stage, the third stage and the fourth stage; The three intermediate feature maps of different sizes (large, medium, and small) input to the cross-scale heterogeneous feature fusion module 1 are: the feature map after convolution of the output of the second stage, the feature map after convolution of the output of the third stage, and the feature map after convolution of the final output of the backbone network; The large, medium, and small intermediate feature maps of different sizes input to the cross-scale heterogeneous feature fusion module 2 are: the feature map after convolution of the output of the first stage, the feature map of the output of the second stage, and the feature map obtained by C3k2 feature extraction and convolution of the output of the cross-scale heterogeneous feature fusion module 1; The head structure includes split head 1, split head 2 and split head 3; The output of the cross-scale heterogeneous feature fusion module 2 is extracted by C3k2 features to obtain a large-scale feature map, which is then output to the segmentation head 1; After the large-scale feature map is processed by the CBS convolution block sequence, it is concatenated with the feature map obtained by C3k2 feature extraction of the output of the cross-scale heterogeneous feature fusion module 1, and then subjected to C3k2 feature extraction again to obtain a medium-sized feature map, which is output to the segmentation head 2; The medium-sized feature map is processed by the CBS convolution block sequence and then concatenated with the final output of the backbone network. It is then extracted by C3k2 features to obtain a small-sized feature map, which is output to the segmentation head three.

8. The method for monitoring water and oil leakage in power scenarios based on instance segmentation according to claim 7 is characterized by: The C3k2 feature extraction means that the input of the C3k2 feature extraction is sequentially processed by a CBS convolution block sequence, split and processed by n C3k feature extractions, the obtained features are concatenated with the split output, and then processed by a CBS convolution block sequence to obtain the output of the C3k2 feature extraction.

9. A readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the method for monitoring water and oil leakage in power scenarios based on instance segmentation according to any one of claims 1 to 8.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the method for monitoring water and oil leakage in electric power scenarios based on instance segmentation according to any one of claims 1 to 8 is implemented.

Images