Mama and frequency domain cooperative oil stain identification method and system under U-shaped structure

Abstract

The invention discloses a Mama and frequency domain cooperative oil stain identification method and system under a U-shaped structure, and relates to the technical field of oil stain leakage identification. The invention aims to solve the problems that the global context sensing capability, the multi-scale adaptability, the calculation efficiency and the detail keeping capability of the existing method are insufficient or are difficult to consider and the like. The technical key points are as follows: a training sample and a test sample are constructed, a Mama and frequency domain collaborative oil stain identification network model under a U-shaped structure is constructed, the model is designed into a U-shaped framework based on a lightweight backbone network SMT, feature redundancy is filtered step by step according to a design concept from top to bottom, the lightweight of the model can be ensured on the basis of high-precision identification, and the identification accuracy is improved. The method achieves the perfect balance of the size and the precision, designs a frequency-guided Mama module, carries out the processing of a high-frequency component and a low-frequency component, and improves the feature operation efficiency through the combination with a Mama assembly. Designing a low-scale convolution perceptron, and training an oil stain identification network model based on Mama and frequency domain cooperation under a U-shaped structure; and performing pixel-level judgment on the oil contamination data by using the learned network model, outputting a black-white two-dimensional segmentation image, and outputting a prediction result.

Description

Technical Field

[0001] This invention relates to the field of oil spill identification technology, specifically to an oil spill identification method and system based on a U-shaped structure and Mamba and frequency domain collaboration. Background Technology

[0002] Oil spills at sea and industrial oil pollution incidents pose serious threats to marine ecosystems, coastal economies, and human health. Therefore, developing rapid and accurate remote sensing monitoring technologies for oil pollution, enabling precise identification and area calculation of oil-polluted regions, is of great significance for environmental governance and disaster emergency response. In recent years, deep learning-based technologies, particularly convolutional neural networks (CNNs) and visual transformers (ViTs), have become research hotspots in this field. For example, the existing technology CN112802006B discloses a deep learning-based edge computing method for identifying oil pollution in motors. This method separates the training, pruning, and inference prediction of the deep learning model. The deep learning model is trained and pruned in the cloud, adapting to the resources and computing power of edge devices. The deep learning model is then used on local edge devices for oil pollution detection, meeting the need for real-time processing of oil pollution identification data. The deep learning model is small and highly accurate. The prior art, document number CN114943895B, discloses a method, device, electronic equipment, and storage medium for identifying oil pollution on the ground in substations. The purpose of this prior art is to improve the accuracy and speed of identifying oil pollution on the ground.

[0003] CNN-based models have achieved remarkable results in image segmentation tasks due to their local connectivity and weight sharing characteristics. However, the inherent local receptive field of CNNs limits their ability to capture long-range spatial dependencies. Oil spills in remote sensing images often exhibit irregular shapes, scattered distribution, and blurred boundaries with the background. This lack of long-range contextual information can easily lead to misjudgment and missed detection of large areas or complex edge oil spills.

[0004] To overcome the limitations of CNNs, the Visual Transformer (ViT) model was introduced, demonstrating superior performance in various visual tasks thanks to its powerful global modeling capabilities based on its self-attention mechanism. However, ViT also has significant drawbacks: First, the computational complexity of its self-attention mechanism is proportional to the square of the image size, resulting in enormous computational and memory overhead when processing high-resolution remote sensing images, making lightweight deployment difficult. Second, standard ViT lacks effective perception of multi-scale information during feature extraction, while oil slicks vary greatly in scale from small oil films to large oil slicks. This insensitivity to scale changes directly affects recognition accuracy, especially for detecting small areas of oil slicks.

[0005] Recently, State Space Models (SSMs), particularly the Mamba architecture, have offered new insights for visual tasks due to their linear computational complexity and global receptive field in long sequence modeling. However, directly applying Mamba to complex oil spill recognition tasks still faces challenges. Simple sequence modeling may not adequately capture high-frequency details in images, such as the sharp boundaries between oil spills and water bodies, and shorelines, which are crucial for accurate segmentation. Frequency domain analysis techniques have proven effective in enhancing the model's ability to extract image texture and edge information, but existing methods rarely deeply integrate Mamba's global context modeling capabilities with frequency domain analysis techniques to collaboratively address the long-range dependencies and fine boundary segmentation problems in oil spill recognition.

[0006] In summary, existing oil spill recognition models suffer from deficiencies or struggle to achieve a balance in terms of global context awareness, multi-scale adaptability, computational efficiency, and detail preservation. Therefore, there is an urgent need in this field for a novel technical solution that can efficiently and accurately identify oil spill regions at different scales, while simultaneously ensuring both lightweight model design and accurate boundary segmentation. Summary of the Invention

[0007] The technical problem to be solved by this invention is:

[0008] The purpose of this invention is to provide a method and system for oil stain identification based on Mamba and frequency domain collaboration under a U-shaped structure, in order to solve the problems of existing oil stain identification models having deficiencies or difficulty in simultaneously achieving global context awareness, multi-scale adaptability, computational efficiency, and detail preservation.

[0009] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0010] The oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure includes the following steps:

[0011] Step 1: Construct training and test samples. Collect images of oil spills generated during oilfield or industrial production processes. If the data volume is insufficient, use data augmentation to expand the data volume. Data preprocessing ensures the diversity and standardization of the data samples. Then, divide the prepared dataset into training and test sets in a 7:3 ratio.

[0012] Step 2: Construct a U-shaped oil pollution identification network model that combines Mamba and frequency domain technologies;

[0013] (1) The model is based on the lightweight backbone network Scale-aware Modulation Transformer (SMT) and designed with a U-shaped architecture. According to the top-down to bottom-up design concept, feature redundancy is filtered step by step, which can ensure the lightweight nature of the model while achieving high-precision recognition, and achieve a perfect balance between size and accuracy.

[0014] (2) To effectively improve the utilization rate of features, we designed a frequency-guided Mamba module from the frequency domain perspective, which processes high-frequency and low-frequency components separately and combines them with Mamba components to improve the efficiency of feature computation. This design enables the model to better explore the dependencies between frequency domains, capture useful frequency domain information, and guide the model to train efficiently.

[0015] (3) To enhance the representational power of low-level features, we designed a low-scale convolutional perceptron during the model training process. As is well known, low-level features retain more detailed information. Focusing on the completeness of feature information as the centrality factor helps the model learn stronger discriminative abilities.

[0016] Step 3: Input the training set and test set of the data in Step 1 into the oil pollution identification network model based on Mamba and frequency domain collaboration under the U-shaped structure, and train the model on the dataset to find the parameter combination that makes the model perform best.

[0017] Step 4: Using the learned network model, perform pixel-level judgment on the oil pollution data and output a black-and-white two-tone segmentation map. Output the prediction results, with the target area being white and the rest of the background being black.

[0018] The specific method for step two of the above scheme is as follows:

[0019] Step 1: Connect to the lightweight backbone network SMT and construct a U-shaped architecture model;

[0020] a. The bottom-up cascade decoder process using convolutional layers with a kernel size of 3×3, batch normalization, and ReLU activation functions stacked into a U-shaped architecture;

[0021] b. For each layer of features extracted from the backbone model, the size of each layer of features in the Cascade Decoder is adjusted to be consistent with the size of the Cascade Decoder using a 1×1 convolutional layer.

[0022] Step 2: Combine Discrete Wavelet Transform (DWT) technology to perform frequency division operation, divide the feature into high-frequency and low-frequency dual branches, and fuse the features under the action of Mamba block;

[0023] a. For high-frequency components, use convolutional layers with kernel sizes of 1×1 and 3×3, and add layer normalization and attention mechanisms during the change process to extract usable features;

[0024] b. Use Mamba blocks to identify target features for low-frequency components;

[0025] c. Use element-wise splicing to fuse the two frequency components into a useful feature containing accurate information.

[0026] Step 3: Design a dedicated low-scale convolutional perceptron (L-ConvA) for low-level features, focusing on enhancing the perception and utilization of low-level features;

[0027] a. Pre-branch structure: First, it undergoes preliminary processing through depthwise convolution and ordinary convolution, and then global average pooling and activation function are used to multiply the features element by element;

[0028] b. Post-branch structure: Consists of identical double branches, with features from both branches concatenated element-wise. A convolutional layer is then used to compress the channels.

[0029] The specific method for step three above is as follows:

[0030] a. Divide the graph dataset transformed in step one into a training set and a test set with a 7:3 ratio. The model performs a normalization operation before reading in the dataset.

[0031] b. The model uses the Adam optimizer as the overall training method, with mean squared error as the loss function, an initial learning rate of 0.0001, and 150 epochs.

[0032] The present invention has the following beneficial effects:

[0033] This invention is based on a lightweight backbone network, the Scale-aware Modulation Transformer (SMT), designed with a U-shaped architecture. Following a top-down to bottom-up design philosophy, it combines Mamba and frequency domain techniques to mine oil spill areas, thereby accurately identifying their location and area. This invention's oil spill identification model balances global context awareness, multi-scale adaptability, computational efficiency, and detail preservation, achieving high accuracy in oil spill identification and localization prediction. Specific advantages are as follows:

[0034] 1. Superior Global Perception and Local Detail Capture Capabilities: This invention innovatively introduces the Fre-Mamba module, a frequency-guided Mamba module. It utilizes frequency domain analysis techniques to transform spatial domain features into the frequency domain for processing, thereby more accurately locating large, irregularly shaped, and dispersed oil slicks. Leveraging its linear computational complexity, the Mamba module can efficiently model the global long-range spatial dependencies of oil slicks in remote sensing images, effectively addressing the limitation of the receptive field in CNNs.

[0035] 2. Refined Boundary Segmentation and Feature Enhancement: A low-scale convolutional perceptron (L-ConvA) is designed. This technique explicitly mines and enhances target components in the image, such as the edges and texture information of oil stains, complementing features extracted from higher-level features in a global context. This synergy makes the model highly sensitive to the blurred boundaries between oil stains and the background, significantly improving the accuracy and continuity of segmentation boundaries, and providing a more reliable data foundation for subsequent area calculations.

[0036] 3. Powerful Multi-Scale Oil Stain Recognition Capability: The lightweight backbone network SMT used in this invention inherently possesses scale awareness capabilities. Combined with the classic top-down and bottom-up multi-scale feature fusion path of the U-shaped architecture, the model can effectively perceive information simultaneously from deep semantic features and shallow detail features. This design enables the invention to adaptively identify targets of different scales, from tiny oil films to large-scale oil stains, effectively overcoming the blind spots of single-scale models in complex scenarios and significantly reducing the false negative rate.

[0037] 4. Balance between high accuracy and high efficiency: The SMT backbone network used in this invention features a lightweight design, while the Mamba module offers computational complexity advantages compared to traditional self-attention mechanisms. Therefore, the entire model maintains high-accuracy recognition performance while effectively controlling the number of parameters and computational overhead, achieving an excellent balance between accuracy and efficiency. This makes the invention easier to deploy on edge devices or emergency response platforms with limited computing resources, meeting the urgent needs of real-time or near-real-time oil spill monitoring. Attached Figure Description

[0038] Figure 1 This is a framework diagram of an oil stain identification model based on a U-shaped structure and the collaboration between Mamba and the frequency domain, according to the present invention.

[0039] Figure 2 This is a schematic diagram showing the performance results of the model on the test set. Detailed Implementation

[0040] The following is in conjunction with the appendix Figure 1-2 The implementation of the oil stain identification method based on Mamba and frequency domain collaboration under the U-shaped structure described in this invention is explained and described as follows:

[0041] This oil spill identification method, based on a U-shaped structure and Mamba and frequency domain collaboration, firstly constructs the graph structure dataset required by the model; secondly, it builds an oil spill identification model based on a U-shaped structure and Mamba and frequency domain collaboration; finally, it inputs the generated graph data into the oil spill identification model based on a U-shaped structure and Mamba and frequency domain collaboration for training, and uses the trained model to identify the target oil spill region and area. Details are as follows:

[0042] Step 1: Construct the graph dataset needed for model training

[0043] This step is used to construct training and test samples, collect images of oil spills generated during oilfield or industrial production, expand the data volume using data augmentation when the data volume is insufficient, ensure the diversity and standardization of data samples through data preprocessing, and then divide the dataset into training and test sets according to the principle that the training set accounts for a larger proportion.

[0044] Images of oil spills from real-world scenarios, including oilfield sites, industrial pipelines, and surveillance videos, are collected. These images should cover diverse conditions and perspectives, including varying lighting, weather, oil type, and background complexity, to ensure data diversity and representativeness. The collected raw images undergo initial screening to remove blurry, duplicate, or irrelevant images, forming the initial image library.

[0045] Step 2: Construct an oil stain identification model that combines Mamba and frequency domain under a U-shaped structure. The specific steps are as follows;

[0046] 1. This model is based on the lightweight backbone network SMT (Scale-aware Modulation Transformer) and is designed with a U-shaped architecture. It filters feature redundancy step by step according to the design concept from top to bottom to top, ensuring the lightweight nature of the model while maintaining high-precision recognition, thus balancing size and accuracy.

[0047] 2. From the perspective of frequency domain, a frequency-guided Mamba module is designed to process high-frequency and low-frequency components separately. Combined with Mamba components, the efficiency of feature computation is improved, enabling the model to explore the dependencies between frequency domains, capture useful frequency domain information, and guide the model to train efficiently.

[0048] 3. During the training process of the model, a low-scale convolutional perceptron is designed to enhance the performance of low-level features. The centrality is placed on the completeness of the feature information, so that the model can learn the ability to distinguish.

[0049] Step 1: A lightweight backbone network (SMT) architecture is adopted, utilizing 1×1 and 3×3 convolutional layers, batch normalization, and ReLU activation to construct a U-shaped architecture. Features are combined element-wise using a cascade decoder. By filtering feature redundancy step-by-step through top-down and bottom-up paths, high segmentation accuracy is maintained while significantly compressing model size and computational cost, making it easy to deploy on industrial embedded devices. This achieves an excellent balance between high accuracy and lightweight design.

[0050] (1)

[0051] Where EB(·) represents the coding block, I represents the image data input to the model, and f i These are the hierarchical features after passing through the backbone network.

[0052] Step 2: Innovatively, a Frequency-Guided Mamba module (Fre-Mamba) is introduced to transform features to the frequency domain and separate high-frequency and low-frequency components. This design enables the model to deeply mine dependencies between frequency domain components, especially enhancing the capture of detailed high-frequency information and structural low-frequency information, thereby significantly improving feature utilization and segmentation boundary accuracy. This achieves superior frequency domain feature utilization capabilities.

[0053] a. Frequency Division Stage: Combining Discrete Wavelet Transform (DWT) technology, under the multi-level decomposition characteristics, the image structure is divided into two different frequency components, high and low. l and f h .

[0054] (2)

[0055] b. Mamba Embedding Stage: As a variant of the Selective State Space Model (SSM), Mamba has outstanding capabilities in capturing and analyzing long-range dependencies and consumes less computational resources compared to the self-attention mechanism in Transformer. Furthermore, Mamba significantly enhances spatial brightness features by selectively modeling low-frequency domain information.

[0056] (3)

[0057] (4)

[0058] (5)

[0059] Where LN(·) represents the LayerNorm operation, DWT(·) represents the DWT function, Mamba(·) represents the Mamba function, split(·) represents the feature segmentation operation, and Conv n (·) represents a convolution operation with an n×n kernel size, and Att(·) represents the attention mechanism, which is spatial attention in this structure. Finally, Cat(·) is used to concatenate the high-frequency and low-frequency feature components element by element.

[0060] Step 3: By designing a low-scale convolutional perceptron (L-ConvA), the focus is on enhancing the perception and utilization of low-level features. These features contain rich texture and detail information. This module ensures that this crucial information is not lost during training, enabling the model to learn more discriminative feature representations and more effectively identify subtle and diffused oil stains. This achieves powerful detail preservation and discrimination capabilities.

[0061] a. Pre-branch structure: First, it undergoes preliminary processing through depthwise convolution and ordinary convolution, and then global average pooling and activation functions are used to multiply the features element by element.

[0062] (6)

[0063] b. Post-branch structure: Consists of identical double branches, with features from both branches concatenated element-wise. A convolutional layer is then used to compress the channels.

[0064] (7)

[0065] Here, GAP(·) represents the global average pooling function, DW. n (·) Depth-wise convolution with a kernel size of n×n (·) represents the Sigmoid function and ⨂ represents element-wise multiplication; the other symbols have the same meaning as in the above formula.

[0066] Step 3: Input the training set and test set of the data divided in Step 1 into the oil pollution identification network model based on Mamba and frequency domain collaboration under a U-shaped structure. Train the model on the dataset to find the parameter combination that makes the model perform best. That is, process the data in Step 1 and input it into Step 2 for training to obtain the optimal parameter combination of the oil pollution identification model based on Mamba and frequency domain collaboration under a U-shaped structure.

[0067] a. Divide the dataset from step one into a 7:3 ratio, and input the resulting training and test sets into the model for training. A composite loss function is used, which integrates the binary cross-entropy (BCE) loss and the intersection-over-union (IoU) loss. The total loss L t The formula is as follows:

[0068] (8)

[0069] here This represents the BCE loss calculation performed on the feature map. This represents the IoU loss calculation performed on the feature map, where P represents the result of predicting the feature map, and G... t Represents the truth graph (GT).

[0070] b. The model uses the Adam optimizer as the overall training method, with mean squared error as the loss function, a learning rate of 0.0001, and 150 epochs. It then converges.

[0071] Step 4: Using the trained network model, perform pixel-level judgment on the oil stain data and output a black-and-white binary segmentation map. The predicted result is output, with the target area represented by white and the rest of the background by black. Randomly input an image to be predicted, test the model using the trained parameter weights, and finally visualize the predicted image.

[0072] The specific method for step one is as follows:

[0073] a. This step is used to build a high-quality, standardized dataset to lay the foundation for subsequent model training and evaluation. Collect oil pollution images from real-world scenarios such as oilfield sites, industrial production pipelines, and surveillance videos. The images should cover various scene conditions and perspectives, as well as different lighting, weather, types of oil pollution, and background complexity to ensure the diversity and representativeness of the data. Perform preliminary screening on the collected raw images to remove blurry, duplicate, or irrelevant images to form the original image library.

[0074] b. To address the issues of insufficient oil spill image data or imbalanced samples, various data augmentation techniques are employed to expand the original image database, thereby improving the model's generalization ability and robustness. These data augmentation methods include, but are not limited to: geometric transformations (random horizontal and vertical flipping, random rotation, random scaling and cropping to simulate different shooting angles); color and brightness perturbations (minor random adjustments to the image's hue, saturation, and brightness to simulate lighting changes and device differences); noise injection (adding Gaussian or salt-and-pepper noise to enhance the model's robustness to image quality degradation); and simulated occlusion (randomly adding rectangular occlusion blocks to simulate potential partial occlusion situations). Through the combined application of these methods, the original dataset is expanded to three times its original size, ensuring sufficient training samples. c. The standardized dataset prepared in the above steps is randomly divided into a training set and a test set in a 7:3 ratio. The training set is used for learning and optimizing model parameters, while the test set serves as independent, unseen data for objectively evaluating the model's final performance and generalization ability. To ensure the fairness of the classification, a stratified sampling strategy is adopted to keep the proportion of each category in the training and test sets consistent with that in the overall dataset.

[0075] Verification of the technical effects of the present invention, such as... Figure 2 The diagram shows the model's performance on the test set. This model was trained using 3345 sample images, and based on the trained weight parameters, it was systematically validated on an independent test set containing 839 images. The test results demonstrate that this method exhibits excellent prediction performance, accurately capturing target contours from local features to global structural levels, effectively anchoring and preserving the target's detailed features. The model output is structurally highly consistent with the ground truth annotations, showing good structural reconstruction ability and boundary recognition accuracy.

[0076] In summary, the method proposed in this invention is based on an identification model built from marine and industrial oil spill production data. Because this invention fully considers the impact of model structure and feature information extraction methods on the identification and location of oil spill targets, the method provided by this invention has high accuracy in identifying oil spill targets. The method proposed in this invention solves the technical problems raised in this invention, and simulation experiments and practical applications have verified the technical effects and practicality claimed by this invention.

[0077] The oil stain identification method (algorithm) proposed in this invention, which combines Mamba and frequency domain collaboration under a U-shaped structure, is the underlying technical core of this invention. Various products can be derived based on this algorithm.

[0078] Based on the method proposed in this invention, a U-shaped structure Mamba and frequency domain collaborative oil stain identification system is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the above-mentioned U-shaped structure Mamba and frequency domain collaborative oil stain identification method during runtime.

[0079] The developed system (software) computer program is stored on a computer-readable storage medium. This computer program is configured to, when called by a processor, implement the steps of the aforementioned oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0080] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0081] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0082] The above description represents only some embodiments of the present invention, but the present invention should not be limited to the content disclosed in these embodiments and accompanying drawings. Therefore, any equivalent or modified versions made without departing from the technical solutions disclosed in the present invention fall within the scope of protection of the present invention.

Claims

1. A method for oil stain identification using Mamba and frequency domain collaboration under a U-shaped structure, characterized in that... Includes the following steps: Step 1: Construct training and test samples: Collect images of oil spills generated during oilfield or industrial production processes. If the amount of data is insufficient, use data augmentation to expand the amount of data. Ensure the diversity and standardization of data samples through data preprocessing. Then, divide the dataset into training set and test set according to the principle that the training set accounts for a larger proportion. Step 2: Construct a U-shaped oil pollution identification network model that combines Mamba and frequency domain technologies; (1) The model is based on the lightweight backbone network SMT and designed as a U-shaped architecture. It filters feature redundancy step by step according to the design concept from top to bottom to top, ensuring the lightweight nature of the model on the basis of high-precision recognition, so as to balance the size and accuracy. (2) From the perspective of frequency domain, a frequency-guided Mamba module is designed to process high-frequency and low-frequency components separately. Combined with Mamba components, the efficiency of feature operation is improved, enabling the model to explore the dependencies between frequency domains, capture useful frequency domain information, and guide the model to train efficiently. (3) Design a low-scale convolutional perceptron during the training process of the model to enhance the performance of low-level features, place centrality on the completeness of feature information, and enable the model to learn discrimination ability. Step 3: Input the training set and test set of the data in Step 1 into the oil pollution identification network model based on Mamba and frequency domain collaboration under the U-shaped structure, and train the model on the dataset to find the parameter combination that makes the model perform best. Step 4: Using the learned network model, perform pixel-level judgment on the oil pollution data and output a black-and-white two-tone segmentation map. Output the prediction results, with the target area being white and the rest of the background being black.

2. The oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure according to claim 1, characterized in that, The specific method for step one is as follows: a. Construct a high-quality, standardized dataset for subsequent model training and evaluation. Collect oil pollution images from real-world scenarios such as oilfield sites, industrial production pipelines, and surveillance videos. The images should cover various scene conditions and perspectives, as well as different lighting, weather, types of oil pollution, and background complexity to ensure the diversity and representativeness of the data. The collected raw images are initially screened to remove blurry, duplicate, or irrelevant images, forming the original image library; b. Due to insufficient oil spill image data or imbalanced samples, various data augmentation techniques are used to augment the original image library to improve the model's generalization ability and robustness. The data augmentation methods used include, but are not limited to, geometric transformations: random horizontal and vertical flipping, random rotation, random scaling and cropping, to simulate different shooting angles. Color and brightness perturbation: Make small random adjustments to the hue, saturation, and brightness of the image to simulate changes in lighting and device differences; Noise injection: Add Gaussian noise or salt and pepper noise to enhance the model's robustness to image quality degradation; Simulated occlusion: Randomly add rectangular occlusion blocks to simulate possible partial occlusion situations in the field; by combining the above methods, the size of the original dataset is expanded to 3 times the original size to ensure sufficient training samples; c. The normalized dataset prepared in the above steps is randomly divided into training set and test set in a 7:3 ratio; the training set is used for learning and optimizing model parameters, while the test set is used as independent, unseen data to objectively evaluate the final performance and generalization ability of the model, so that the proportion of each category in the training set and test set is consistent with the overall dataset.

3. The oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure according to claim 1 or 2, characterized in that, The specific method for step two is as follows: Step 1: A lightweight backbone network (SMT) architecture is adopted, utilizing 1×1 and 3×3 convolutional layers, batch normalization, and ReLU activation to construct a U-shaped architecture. Features are combined element-wise using a cascaded decoder. Redundancy is filtered step-by-step through top-down and bottom-up paths to ensure high segmentation accuracy while compressing model size and computational cost, making it easy to deploy on industrial embedded devices and achieving an excellent balance between high accuracy and lightweight design. （1） Where EB(·) represents the coding block, I represents the image data input to the model, and f i These are the hierarchical features after passing through the backbone network; Step 2: Introduce the Frequency-Guided Mamba module, i.e., Fre-Mamba, to transform features to the frequency domain and separate high-frequency and low-frequency components; this enables the model to deeply explore the dependencies between frequency domains and enhances the capture of detailed high-frequency and structural low-frequency information. a. Frequency Division Stage: Combining Discrete Wavelet Transform (DWT), under the multi-level decomposition characteristics, the image structure is divided into two different frequency components, high and low. l and f h , （2） b. Mamba Embedding Stage: As a variant of the Selective State-Space Model (SSM), Mamba enhances spatial brightness features by selectively modeling low-frequency domain information, based on its long-range dependency capture and analysis capabilities and low computational resource consumption. （3） （4） （5） Where LN(·) represents the LayerNorm operation, DWT(·) represents the DWT function, Mamba(·) represents the Mamba function, split(·) represents the feature segmentation operation, and Conv n (·) is a convolution operation with an n×n kernel size, Att(·) represents the attention mechanism, and the structure uses a spatial attention mechanism. Finally, Cat(·) is used to concatenate the high-frequency and low-frequency component features element by element. Step 3: Enhance the perception and utilization of low-level features by designing a low-scale convolutional perceptron L-ConvA. The underlying features contain texture and detail information. This module ensures that this crucial information is not lost during training, enabling the model to learn more discriminative feature representations for recognizing fine and diffused oil stains, thus achieving detail preservation and discriminative capabilities. a. Pre-branch structure: First, it undergoes preliminary processing through depthwise convolution and ordinary convolution, and then global average pooling and activation function are used to multiply the features element by element; （6） b. Post-branch structure: It consists of identical double branches. The features on the two branches are spliced ​​together element by element, and then a convolutional layer is used to compress the channels. （7） In the formula, GAP(·) represents the global average pooling function, and DW n (·) Depth-wise convolution with a kernel size of n×n (·) represents the Sigmoid function and the element-wise multiplication of ⨂.

4. The oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure according to claim 1, characterized in that, The specific method for step three is as follows: a. Divide the dataset from step one into a 7:3 ratio, and input the divided training and test sets into the model for training. Use a composite loss function that integrates the binary cross-entropy (BCE) loss and the intersection-over-union (IoU) loss; the total loss L t The formula is as follows: （8） In the formula This represents the BCE loss calculation performed on the feature map. This represents the IoU loss calculation performed on the feature map, where P represents the result of predicting the feature map, and G... t Represents the truth graph (GT); b. The model uses the Adam optimizer as the overall training method, with mean squared error as the loss function, a learning rate of 0.0001, and 150 epochs. It then converges.

5. The oil stain identification method based on Mamba and frequency domain collaboration under a U-shaped structure according to claim 1, characterized in that, The specific method for step four is as follows: A random image data to be predicted is input, the model is tested using trained parameter weights, and finally the predicted image is visualized.

6. An oil stain identification system based on Mamba and frequency domain collaboration under a U-shaped structure, characterized in that: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 5, and executes the steps of the oil stain identification method of Mamba and frequency domain collaboration under a U-shaped structure when running.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the oil stain identification method based on Mamba and frequency domain coordination under a U-shaped structure according to any one of claims 1 to 5.

8. An oil stain identification device with Mamba and frequency domain coordination under a U-shaped structure, characterized in that: The device includes At least one processor is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor. When the at least one processor executes the instructions, it enables the at least one processor to perform the oil stain identification method of Mamba and frequency domain coordination under a U-shaped structure according to any one of claims 1-5, thereby realizing oil stain identification.

Images