Oil spill detection method based on HSI color space instance segmentation

Abstract

The application discloses an oil stain leakage detection method based on HSI color space instance segmentation, collects an oil stain leakage detection image of an oil-filled equipment, and then performs HSI color space conversion; an oil stain leakage detection network of the oil-filled equipment based on instance segmentation is constructed and trained, and then the oil stain leakage detection network of the oil-filled equipment based on instance segmentation after training is used to detect the HSI color space oil stain leakage detection image of the oil-filled equipment; the oil stain leakage detection network of the oil-filled equipment based on instance segmentation comprises a multi-scale candidate frame extraction network and a detection segmentation head network, the multi-scale candidate frame extraction network performs feature extraction on the HSI color space oil stain leakage detection image of the oil-filled equipment, and the detection segmentation head network comprises an oil-filled equipment detection branch and an oil stain segmentation branch. In the HSI color space, after multi-scale extraction, the area of the oil-filled equipment and the oil stain area are respectively judged, the oil stain area ratio is calculated to judge the oil stain leakage degree of the oil-filled equipment, and the detection effect is improved.

Description

Technical Field

[0001] This invention belongs to the field of substation detection technology, specifically relating to an oil leakage detection method based on HSI color space instance segmentation. Background Technology

[0002] Substations are fundamental nodes in the power grid. To ensure their stable operation, it is essential to maintain the normal operating condition of their equipment. Under current technological and economic conditions, many critical components in substations, such as transformers, instrument transformers, and capacitors, are still oil-filled. Oil leaks can occur during production, transportation, operation, or maintenance, affecting not only the cleanliness of the equipment but also potentially triggering alarms or even fires and explosions due to reduced oil levels. Therefore, detecting oil leaks in oil-filled equipment within substations is crucial.

[0003] Currently, oil leak detection relies heavily on manual inspections, and the judgment of oil leaks depends primarily on human experience, inevitably leading to oversights. Furthermore, a single work team often needs to inspect multiple substations, resulting in long inspection cycles and difficulty in timely detection of early problems. To shorten the production safety inspection cycle, numerous video surveillance systems have been widely deployed. However, relying on manual observation of multiple video surveillance channels is not only inefficient, but achieving 24 / 7 oil leak monitoring also requires a significant amount of manpower. Moreover, manually judging oil leaks based on video information is more difficult than on-site inspection, and is limited by the display quality of the monitors.

[0004] The development of production safety assurance is undergoing a transformation from "safety monitoring" to "intelligent monitoring." Utilizing artificial intelligence to analyze monitoring videos enables rapid analysis of large amounts of data, significantly shortening inspection cycles. However, current methods often directly process images or videos of oil leaks. Substations, with their complex environments, often present oil leaks that are not readily apparent against the background. Furthermore, shadows caused by lighting and obstructions can interfere with oil stain detection. To make leaks more visible, some methods employ fluorescence detection, but this requires additional deployment of ultraviolet fluorescent lamps and cameras capable of capturing fluorescence, resulting in high deployment costs.

[0005] With the development of artificial intelligence and computer vision technologies, object detection methods based on deep convolutional neural networks have achieved initial applications in the field of smart power grid inspection. However, deep learning-based inspection methods require a large amount of data to train the model. To apply them to tasks with limited data, engineers often initialize some network parameters using pre-trained models trained on large amounts of data and then optimize them on a small amount of data. For example, the pre-trained models of mainstream object detection models are currently trained on the ImageNet image classification dataset. However, ImageNet training is performed in the RGB color space, aiming to better distinguish various targets in real life. For oil leakage tasks, the color difference between normal and healthy oil-filled equipment and leaking oil-filled equipment is difficult to describe in the RGB color space. Therefore, detection methods built using the RGB color space cannot achieve good detection results for oil leakage. Summary of the Invention

[0006] Existing AI-based methods for detecting oil leaks in substation oil-filled equipment operate in the RGB color space. However, the color changes caused by oil leaks to different colored oil-filled equipment are subtle, resulting in poor detection performance in the RGB color space. Furthermore, the distribution of oil contaminants on the equipment lacks a clear pattern, making it difficult for bounding box-level target detection methods to differentiate between normal and leaking equipment based on this irregular distribution. Existing methods require a large amount of data on oil-filled equipment with leaks, which is difficult to obtain in practice. To overcome these shortcomings, this invention provides an oil leak detection method based on HSI color space instance segmentation. This method converts the RGB color space to the HSI color space to expand the color features of oil leaks (HSI represents hue, saturation, and intensity, respectively), and pre-trains an instance-segmentation-based oil leak detection network for oil-filled equipment using the ImageNet dataset in the HSI color space, thus improving detection performance.

[0007] This invention is achieved through the following technical solution: an oil leakage detection method based on HSI color space instance segmentation. The method involves acquiring oil leakage detection images of oil-filled equipment, then performing HSI color space conversion to obtain converted HSI color space oil leakage detection images of the oil-filled equipment; constructing and training an instance segmentation-based oil leakage detection network for oil-filled equipment; and then using the trained instance segmentation-based oil leakage detection network to detect oil leakage in the HSI color space oil-filled equipment images. The instance segmentation-based oil leakage detection network includes multi-scale weather... The system employs a bounding box extraction network and a detection and segmentation head network. The multi-scale candidate bounding box extraction network extracts features from oil leakage detection images of oil-filled equipment in the HSI color space. The detection and segmentation head network consists of two branches: an oil-filled equipment detection branch and an oil leakage segmentation branch. The oil-filled equipment detection branch is responsible for selecting the region of the oil-filled equipment from the complex background. The oil leakage region segmentation branch is responsible for performing pixel-level oil leakage classification within the region of the oil-filled equipment predicted by the oil-filled equipment detection branch. The degree of oil leakage is determined based on the proportion of the oil leakage region segmented by the oil leakage segmentation branch to the total area of ​​the oil-filled equipment.

[0008] Further optimization reveals that the HSI color space divides an image into three components: hue, saturation, and intensity. The formula for obtaining the hue value H from the RGB color space is:

[0009]

[0010] in:

[0011]

[0012] The formula for converting the saturation value S is:

[0013]

[0014] The conversion formula for intensity value I is:

[0015]

[0016] In the formula, R is the red intensity value, G is the green intensity value, B is the blue intensity value, and θ is the angle value.

[0017] Further optimized, the multi-scale candidate box extraction network consists of a backbone network ResNet, a feature pyramid network, a candidate box network, and a RoI Align network. The backbone network ResNet extracts features in layers: a first layer C1, a second layer C2, a third layer C3, a fourth layer C4, and a fifth layer C5. The input to the feature pyramid is the second layer C2, the third layer C3, the fourth layer C4, and the fifth layer C5 of the backbone network. A 1×1 convolutional layer is used to obtain the fifth fusion feature M5, which is then upsampled. The upsampled feature is then added to C4 after the 1×1 convolutional layer to obtain the fourth fusion feature M4. The same operation is then used to obtain the third fusion feature M3 and the second fusion feature M2. To fuse features of different scales together, a 3×3 convolutional layer is used to fuse the different scale features of fusion features M2-M5, obtaining the second feature map P2, the third feature map P3, the fourth feature map P4, and the fifth feature map P5.

[0018] The candidate box network and the RoI Align network collaborate to predict target candidate regions and extract target region features in the image. The candidate box network densely predicts candidate boxes at each feature layer of each scale. After non-maximum suppression post-processing, a certain number of candidate boxes are selected as target regions from each feature layer based on prediction confidence. Based on the positional information of the candidate boxes, more accurate features are obtained using bilinear interpolation of the RoI Align network.

[0019] Further optimization involves using the 7×7×256 features provided by the RoI Align network for prediction in the oil-filled device detection branch. Two fully connected layers are then used for feature integration, and finally, two fully connected layers are used separately to predict the classification and regression vectors. The classification vector of the oil-filled device detection branch has a dimension of 2, representing the oil-filled device and the background, while the regression vector has a dimension of 4, representing the center coordinates, length, and width. For a candidate bounding box containing the target, the classification task uses the cross-entropy loss function, and the regression task uses the smooth L1 loss function.

[0020]

[0021] Where q and p represent the true class vector and predicted vector of the target, respectively, and b t and b p These represent the target's true position vector and predicted vector, respectively.

[0022] Further optimization revealed that the candidate bounding box feature sampling size used in the oil spill segmentation branch is 14×14. The oil spill segmentation branch extracts oil spill leakage features through four 3×3 convolutional layers, and finally obtains a 28×28 classification result through a deconvolutional layer. The classification vector at each location has a dimension of 2, representing the oil spill area and the normal area respectively. For a candidate bounding box, the optimized loss function of the oil spill segmentation branch is:

[0023]

[0024] Where t i and x i These represent the true class vector and the predicted vector for each pixel, respectively.

[0025] The present invention also provides a non-volatile computer storage medium storing computer-executable instructions that can execute the oil leakage detection method based on HSI color space instance segmentation in any of the above embodiments.

[0026] The present invention also provides a computer program product, which includes a computer program stored on a non-volatile computer storage medium. The computer program includes program instructions, which, when executed by a computer, cause the computer to perform the oil leakage detection method based on HSI color space instance segmentation of the above embodiments.

[0027] The beneficial effects of this invention are as follows: The RGB color space is converted to the HSI color space to expand the color features of oil leakage (HSI represents hue, saturation, and intensity, respectively). An instance-segmentation-based oil leakage detection network for oil-filled devices is trained using the ImageNet dataset in the HSI color space. The detection and segmentation head network includes two branches: an oil-filled device detection branch and an oil leakage segmentation branch. The oil-filled device detection branch is responsible for selecting the region of the oil-filled device from a complex background; the oil leakage segmentation branch is responsible for pixel-level oil leakage classification within the small region predicted by the detection branch. Targets where the oil leakage region occupies a certain proportion of the oil-filled device are defined as oil-filled devices with oil leakage. The training of the oil-filled device detection branch does not require data support for oil leakage and can be trained using a large number of normal devices; the oil leakage segmentation branch only needs to segment the oil leakage region in a local area of ​​the image—a device slice—reducing the interference of the background on oil leakage detection. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a schematic diagram of an oil leakage detection network for oil-filled equipment based on instance segmentation.

[0030] Figure 3This is a schematic diagram of multi-scale feature extraction. Detailed Implementation

[0031] The present invention will be further explained in detail below with reference to the accompanying drawings and embodiments.

[0032] Reference Figure 1 This paper presents an oil leakage detection method based on HSI color space instance segmentation. The method involves acquiring oil leakage detection images of oil-filled equipment, performing HSI color space conversion to obtain converted HSI color space oil leakage detection images, constructing and training an instance segmentation-based oil leakage detection network, and then using the trained network to detect oil leakage in the HSI color space oil-filled equipment images. The instance segmentation-based oil leakage detection network includes a multi-scale candidate box extraction network and a detection segmentation head network. The network (multi-scale candidate box extraction network) extracts features from oil leakage detection images of oil-filled equipment in the HSI color space. The detection and segmentation head network includes two branches: an oil-filled equipment detection branch and an oil leakage segmentation branch. The oil-filled equipment detection branch is responsible for selecting the region of the oil-filled equipment from the complex background. The oil leakage region segmentation branch is responsible for performing pixel-level oil leakage classification within the region of the oil-filled equipment predicted by the oil-filled equipment detection branch. The degree of oil leakage of the oil-filled equipment is determined based on the proportion of the oil leakage region segmented by the oil leakage segmentation branch to the area of ​​the oil-filled equipment.

[0033] Natural image instance segmentation networks are typically trained using the RGB color space. The three channels of the RGB color space represent the intensity values ​​of red, green, and blue, respectively, and these values ​​change significantly with varying light intensity. In contrast, the HSI color space divides an image into three components: hue (H), saturation (S), and intensity (I). The hue and saturation values ​​reflect color characteristics unaffected by intensity. The formula for obtaining the hue value H from the RGB color space is:

[0034]

[0035] in:

[0036]

[0037] The formula for converting the saturation value S is:

[0038]

[0039] The conversion formula for intensity value I is:

[0040]

[0041] In the formula, R is the red intensity value, G is the green intensity value, B is the blue intensity value, and θ is the angle value.

[0042] Reference Figure 1 , Figure 2 In this embodiment, the multi-scale candidate box extraction network consists of a backbone network (ResNet), a Feature Pyramid Network (FPN), a Region Proposal Network (RPN), and a RoI Align network. Its purpose is to extract the location of the target region, also known as the candidate box, from the original image; this process requires no manual intervention. Since the distance between the monitoring equipment and different oil-filled devices varies, the size of the oil-filled devices in the image also fluctuates significantly. Therefore, this embodiment uses a multi-scale candidate box network (FPN+RPN). To enable subsequent detection branches to identify whether the target in the candidate box is an oil-filled device, and to segmentation branches to predict the area of ​​oil leakage, the model uses the RoI Align module to extract multi-scale candidate box features based on the candidate box's location. The backbone network uses a residual network, ResNet50, which has deeper layers and better feature extraction capabilities. ResNet50 uses a jumper structure to introduce residual modules, allowing gradients to propagate back along the jumper structure, thereby optimizing the model. ResNet50 is pre-trained on the ImageNet dataset in the HSI color space using a classification task, providing initialization parameters for subsequent detection and segmentation tasks.

[0043] Because the imaging distance range is large and the target scale range in the field of view is large during UAV inspection, the detection effect of a single scale is poor. Therefore, a feature pyramid network is introduced to obtain multi-scale features. The structure diagram is as follows. Figure 3 As shown, the input to the feature pyramid consists of the second-layer features C2, the third-layer features C3, the fourth-layer features C4, and the fifth-layer features C5 of the backbone network. The spatial resolution of these four layers decreases sequentially, while the semantic information contained within the features increases sequentially. The semantic information in shallow layers is relatively poor, and the spatial resolution of deeper features is low, requiring upsampling before fusion with shallower features. Specifically, a 1×1 convolutional layer is first used to obtain the fifth fusion feature M5 and perform upsampling. Then, the upsampled feature is added to C4 after passing through a 1×1 convolutional layer to obtain the fourth fusion feature M4. The same operation is then used to obtain the third fusion feature M3 and the second fusion feature M2. To fuse features of different scales, a 3×3 convolutional layer is used to fuse the different scale features of fusion features M2-M5, obtaining the second feature map P2, the third feature map P3, the fourth feature map P4, and the fifth feature map P5.

[0044] The Candidate Box Network (RPN) and the RoI Align Network collaborate to predict candidate regions of interest (CGIs) and extract features from these regions in an image. The RPN densely predicts candidate boxes at each feature layer of each scale. After non-maximum suppression post-processing, a certain number of candidate boxes are selected as CGIs from each feature layer based on prediction confidence. Based on the positional information of the candidate boxes, the RoI Align Network's bilinear interpolation is used to obtain more accurate features, ensuring that the detection and segmentation head network performs accurate detection and localization.

[0045] In this implementation, the detection and segmentation head network includes two branches: an oil-filled device detection branch and an oil contamination segmentation branch. The oil-filled device detection branch is responsible for selecting the region of oil-filled device from a complex background; the oil contamination segmentation branch is responsible for pixel-level oil contamination classification within the region of oil-filled device predicted by the oil-filled device detection branch. Targets where the oil contamination area occupies a certain proportion of the oil-filled device are defined as oil-filled devices with oil leakage. The oil-filled device detection branch uses 7×7×256 features provided by the RoIAlign network (pooling) for prediction, then uses two fully connected layers for feature integration, and finally uses two fully connected layers to predict the classification vector and regression vector respectively. The classification vector of the oil-filled device detection branch has a dimension of 2, representing the oil-filled device and the background, while the regression vector has a dimension of 4, representing the center coordinates, length, and width. For a candidate box containing the target, the classification task uses the cross-entropy loss function, and the regression task uses the smooth L1 loss function.

[0046]

[0047] Where q and p represent the true class vector and predicted vector of the target, respectively, and b t and b p These represent the target's true position vector and predicted vector, respectively.

[0048] To achieve pixel-level segmentation, high spatial resolution features are required. The oil spill segmentation branch uses candidate bounding box features with a sampling size of 14×14. The oil spill segmentation branch extracts oil spill features through four 3×3 convolutional layers, and finally obtains a 28×28 classification result through a deconvolutional layer. The classification vector at each location has a dimension of 2, representing the oil spill area and the normal area respectively. For a candidate bounding box, the optimized loss function of the oil spill segmentation branch is:

[0049]

[0050] Where t i and x i Let represent the true class vector and the predicted class vector for each pixel, respectively. The loss functions for the detection segmentation head network are:

[0051] L = L mask +L cls +L box

[0052] For a specific area of ​​an oil-filled device, when the oil-stained area occupies more than 5% of the area of ​​the oil-filled device, it is considered an oil-leaking device. The larger the area occupied by the oil-stained area, the greater the degree of oil leakage.

[0053] In this embodiment, the instance-segmentation-based oil leakage detection network for oil-filled equipment involves manually labeling the regions of the oil-filled equipment and the oil-contaminated areas in the image dataset samples during training. The labeled image dataset is then expanded and divided into training and testing sets. The training set is used to train the instance-segmentation-based oil leakage detection network, and the testing set is used to test the accuracy of the trained network. Notably, the training of the oil-filled equipment detection branch does not require data support for oil leakage detection and can utilize a large number of normal devices for training. The oil leakage segmentation branch only needs to segment the oil-contaminated areas within a local area of ​​the image—a slice of the equipment—reducing the interference of the background on oil leakage detection.

[0054] In other embodiments, a non-volatile computer storage medium is provided, which stores computer-executable instructions that can execute the oil leakage detection method based on HSI color space instance segmentation in any of the above embodiments.

[0055] This embodiment also provides a computer program product, which includes a computer program stored on a non-volatile computer storage medium. The computer program includes program instructions, which, when executed by a computer, cause the computer to perform the oil leakage detection method based on HSI color space instance segmentation described above.

[0056] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0057] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0058] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0059] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0060] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0061] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An oil spill detection method based on HSI color space instance segmentation, characterized in that, An oil pollution leakage detection image of the oil-filled equipment is collected, and then HSI color space conversion is performed to obtain an oil pollution leakage detection image of the oil-filled equipment in the converted HSI color space; an oil pollution leakage detection network based on instance segmentation is constructed and trained, and then the trained oil pollution leakage detection network based on instance segmentation is used to detect the oil pollution leakage detection image of the oil-filled equipment in the HSI color space; the oil pollution leakage detection network based on instance segmentation includes a multi-scale candidate box extraction network and a detection segmentation head network; the multi-scale candidate box extraction network extracts features from the oil pollution leakage detection image of the oil-filled equipment in the HSI color space; the detection segmentation head network includes two branches, i.e., an oil-filled equipment detection branch and an oil pollution segmentation branch; the oil-filled equipment detection branch is responsible for selecting the area of the oil-filled equipment from a complex background; the oil pollution segmentation branch is responsible for oil pollution classification at a pixel level in the area of the oil-filled equipment predicted by the oil-filled equipment detection branch, and the degree of oil pollution leakage of the oil-filled equipment is determined according to the proportion of the area of the oil pollution region segmented by the oil pollution segmentation branch to the area of the oil-filled equipment. The multi-scale candidate frame extraction network is composed of a backbone network ResNet, a feature pyramid network, a candidate frame network and a RoIAlign network, the backbone network ResNet is used for hierarchical extraction to obtain a first layer feature C1, a second layer feature C2, a third layer feature C3, a fourth layer feature C4 and a fifth layer feature C5, the input of the feature pyramid is the second layer feature C2, the third layer feature C3, the fourth layer feature C4 and the fifth layer feature C5 of the backbone network, 1 1 convolutional layers are used to obtain the fifth fusion feature M5 and complete the upsampling operation, and then the up-sampled feature is added to the C4 of the 1 1 convolutional layer to obtain the fourth fusion feature M4, and the same operation is used to obtain the third fusion feature M3 and the second fusion feature M2; in order to fuse the features of different scales together, 3 3 convolutional layers are used to fuse the features of different scales of the fusion features M2-M5 to obtain a second feature map P2, a third feature map P3, a fourth feature map P4 and a fifth feature map P5.

2. The oil spill detection method based on HSI color space instance segmentation according to claim 1, characterized in that, The HSI color space divides an image into three components, i.e., hue, saturation and intensity; the formula for obtaining a hue value H from an RGB color space is as follows: ； wherein, ； the conversion formula for a saturation value S is as follows: ； the conversion formula for an intensity value I is as follows: ； where R is the red intensity value, G is the green intensity value, B is the blue intensity value, is an angle value.

3. The oil spill detection method based on HSI color space instance segmentation according to claim 1, characterized in that, The candidate box network and the RoI Align network cooperate to complete prediction of a target candidate region in an image and extraction of a target region feature; the candidate box network densely predicts candidate boxes on each scale of a feature layer; after non-maximum suppression post-processing, a certain number of candidate boxes are selected as target regions from each scale of the feature layer according to a prediction vector; based on position information of the candidate boxes, the RoI Align network is used to obtain more accurate features through bilinear interpolation.

4. The oil spill detection method based on HSI color space instance segmentation according to claim 1, characterized in that, The oil filling equipment detection branch uses the 7 7 256 features to make a prediction, and then uses two fully connected layers to integrate the features, and finally uses two fully connected layers to predict a classification vector and a regression vector respectively; the classification vector of the oil filling equipment detection branch has a dimension of 2, representing the oil filling equipment and the background, and the regression vector has a dimension of 4, representing the center coordinates, length, width and four variables respectively; for a candidate frame containing a target, the classification task uses a cross-entropy loss function, and the regression uses a smooth L1 loss function: ； wherein and Y and Y' represent the true and predicted vectors of the target's class, respectively, and X and X' represent the true and predicted vectors of the target's position, respectively.

5. The oil spill detection method based on HSI color space instance segmentation according to claim 1, characterized in that, The candidate box feature sampling size used by the oil stain segmentation branch is 14 14; the oil stain segmentation branch extracts oil stain leakage features through a 4-layer 3 3-size convolution layer, and finally obtains a classification result of 28 28 size through a deconvolution layer. The classification vector dimension of each position is 2, representing the oil stain area and the normal area respectively. For a candidate box, the optimization loss function of the oil stain segmentation branch is: ； where and represent the true and predicted vectors of classes for each pixel, respectively.

6. A non-transitory computer storage medium storing computer-executable instructions, the computer-executable instructions comprising instructions for: The computer executable instructions can execute the oil pollution leakage detection method based on HSI color space instance segmentation according to any one of claims 1-5.

7. A computer program product, the computer program product comprising a computer program stored on a non-transitory computer storage medium, characterized in that, The computer program includes program instructions, which, when executed by a computer, cause the computer to execute the oil pollution leakage detection method based on HSI color space instance segmentation according to any one of claims 1-5.

Images