Marine organism intelligent detection system based on multi-scale convolution fusion and YOLOv8

Abstract

The invention relates to the technical field of computer vision, in particular to a marine organism intelligent detection system based on multi-scale convolution fusion and YOLOv8, and the system comprises an acquisition module which is used for obtaining the original image data of underwater marine organisms, and processing the original image data to obtain a marine organism data set; the detection module is used for identifying the marine organism data set; through the VM-Unet improved multi-scale underwater image enhancement network and the deeply optimized YOLOv8 algorithm, lightweight and efficient calculation, multi-scale feature fusion and complex scene robust recognition are realized, the precision and real-time performance of underwater image processing and target detection are remarkably improved, and the method has the advantages of being high in robustness and high in robustness. And a technical support with high adaptability and high efficiency is provided for a marine organism intelligent detection system.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8. Background Technology

[0002] With the deepening of the national marine strategy, the demand for intelligent technologies in scenarios such as marine environmental monitoring, underwater infrastructure inspection, and aquaculture monitoring is becoming increasingly urgent. Traditional manual fishing methods are not only inefficient and risky, but also negatively impact the marine ecosystem due to the disposal of fishing gear and other harmful waste. Therefore, replacing traditional manual operations with advanced technologies has become an inevitable trend. The rapid development of deep learning target detection technology has made it possible to build intelligent marine organism detection systems, aiming to achieve efficient and accurate detection of marine organisms through technological innovation and promote the intelligent upgrading of the marine industry. In existing technologies, several patents have been dedicated to solving the problem of underwater target detection. For example, publication number CN118968213A proposes an underwater target recognition method that combines adaptive image enhancement and lightweight detection. It dynamically generates image preprocessing parameters through an HPNN network to optimize the low contrast and color distortion problems of underwater images, and replaces the convolutional modules of the YOLOv8 backbone network with depthwise separable convolutional modules to reduce the number of model parameters. However, this method relies on the first training dataset to learn the filtering parameters. If the training data does not cover extreme scenarios, the dynamic parameters may become invalid, leading to unstable enhancement results. Another patent, CN120047810A, proposes a lightweight adaptive underwater target detection method. By embedding a multi-scale expanded attention mechanism and a novel module design into the backbone network, it significantly enhances the ability to recognize small targets. Simultaneously, it introduces a grouped shuffling convolutional module and a cross-level local network module into the neck network to construct a lightweight structure and reduce computational complexity. However, this method is highly dependent on the distribution characteristics of specific datasets, and its performance may significantly degrade under extreme class imbalances and different water quality environments. Furthermore, patent CN120032118A proposes an adaptive multi-scale feature fusion underwater target detection method. It improves the accuracy of small target detection by optimizing the feature transfer path through dynamic weight allocation and a hybrid loss function. However, this method suffers from excessive computational burden, insufficient training stability, and limited cross-modal generalization ability and redundant computation due to its highly customized structure, ultimately restricting its efficient deployment on edge devices and its ability to migrate to real-world scenarios. Summary of the Invention

[0003] In view of this, in order to address the shortcomings of the prior art, the present invention proposes a marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8, aiming to solve at least one of the problems mentioned in the background art.

[0004] This invention provides a marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8, comprising: an acquisition module for acquiring raw image data of underwater marine organisms and processing the raw image data to obtain a marine organism dataset;

[0005] The detection module is used to identify the marine organism dataset.

[0006] In some embodiments, when the acquisition module processes the raw image data to obtain a marine life dataset, it includes:

[0007] The acquisition module is configured to perform image scaling, random rotation, image blurring, brightness adjustment, mirror flipping, and noise processing on the raw image data.

[0008] In some embodiments, when the detection module is used to identify the marine organism dataset, it includes:

[0009] The detection module is configured to identify the marine organism dataset using the VM-Unet model, which includes a PVM layer, a patch embedding layer, an encoder, a decoder, a final projection layer, and skip connections.

[0010] In some embodiments, the patch embedding layer divides the input image into non-overlapping blocks and maps them to a high-dimensional space with C channels;

[0011] The encoder consists of four stages, with the number of channels gradually increasing in the order of [C, 2C, 4C, 8C]. Each stage contains a VSS module, which reduces the spatial dimension through the PatchMerging operation.

[0012] The PVM layer divides feature X into four sub-features after normalization by the LayerNorm layer. Each sub-feature is input into Mamba for processing and then optimized through residual connections and adjustment factors. Finally, the sub-features are merged and output. The number of channels for each sub-feature is C / 4.

[0013] In some embodiments, the detection module is further configured to extract image features of the marine biological dataset through convolutional layers with kernel sizes of 1, 3, 5, and 7, wherein the number of output channels is 16 in each layer. The image features are subjected to two layers of convolution, normalization, and activation processing, and the dimension is compressed to C / 16*H*W before being multiplied element-wise with the original image features and superimposed on the channel dimension.

[0014] In some embodiments, the detection module is further configured to identify the marine organism dataset using an optimized YOLOv8 target detection algorithm.

[0015] In some embodiments, the detection module is further configured to identify the marine organism dataset using an optimized YOLOv8 object detection algorithm, including:

[0016] A DSC-GAM module is added to the backbone network of YOLOv8. The DSC-GAM module replaces the fully connected layer in the channel attention submodule with depthwise separable convolution to enhance feature extraction and suppress irrelevant information. SCConv is used instead of ordinary convolution in the neck network to reduce computational complexity, and VariFocalLoss is used to optimize the classification loss.

[0017] In some embodiments, adding a DSC-GAM module to the YOLOv8 backbone network includes:

[0018] The input feature F1 is processed by a channel attention mechanism to obtain the intermediate state F2. The channel attention mechanism adjusts the channel weights through two MLP layers and depthwise separable convolution.

[0019] The intermediate state F2 is optimized by a spatial attention mechanism to obtain the output feature F3, which highlights key features through max pooling and average pooling.

[0020] In some embodiments, the detection module is further configured to employ an AdamW optimizer and a cosine annealing learning rate scheduling strategy, with an initial learning rate of 4×10⁻⁶. -3 Attenuation to 1×10 -6 The training batch size was 16, and the training period was 500 epochs to train the YOLOv8 model.

[0021] Compared with existing technologies, the beneficial effects of this invention are as follows: The proposed multi-scale underwater image enhancement network based on VM-Unet significantly improves underwater image processing performance through innovative design. Its core advantages are: First, by introducing an asymmetric encoder-decoder structure and combining it with a parallel visual Mamba layer (PVM), the number of parameters is reduced by 93.6% while retaining global feature modeling capabilities, achieving lightweight and efficient computation; second, a multi-scale feature convolutional fusion module (MFCFM) based on an attention mechanism is designed to effectively integrate underwater image features at different scales, enhancing the recovery capabilities for color distortion, low contrast, and blurred details; finally, a joint optimization strategy using Charbonnier loss and edge loss is employed to improve the naturalness of image colors while preserving key edge information. Experiments show that this method outperforms mainstream algorithms in objective metrics such as PSNR and SSIM, as well as subjective evaluations, requiring only 0.2 seconds for single-image processing. It significantly improves the accuracy of downstream tasks such as image segmentation and target recognition for underwater robots, exhibiting both strong robustness and real-time performance. The technical advantages of this invention in image recognition are mainly reflected in the following aspects: By deeply optimizing the YOLOv8 algorithm, the detection performance in complex underwater scenes is significantly improved. Specifically, a lightweight DSC-GAM module (based on a global attention mechanism using depthwise separable convolution) is innovatively embedded in the backbone network, enhancing the ability to filter key target features and suppressing background interference; SCConv is used in the neck network instead of standard convolution, significantly reducing computational complexity; and the VariFocal Loss function is introduced to effectively alleviate the problem of imbalanced underwater sample classes, especially improving the recognition sensitivity of difficult-to-classify positive samples (such as sea cucumbers and small sea urchins whose colors are similar to the background). Experimental verification shows that the improved algorithm significantly reduces the false negative rate compared to the original YOLOv8 in scenes with severe color cast, target occlusion, and dense small targets, while maintaining high robustness, providing a precise and efficient recognition core for intelligent marine organism detection systems.

[0022] The above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0023] Other features and aspects of this disclosure will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 Functional block diagram of a marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 provided in an embodiment of the present invention;

[0026] Figure 2 The original underwater image and its corresponding high-quality comparison image are provided by the marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 in the embodiments of the present invention.

[0027] Figure 3 A comparison of results of different enhancement methods for the marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 provided in the embodiments of the present invention;

[0028] Figure 4 Line graphs showing the average UIQM and UCIQE for each method in the intelligent marine biological detection system based on multi-scale convolutional fusion and YOLOv8 provided in the embodiments of the present invention.

[0029] Figure 5 The image shows the test results of the improved algorithm model of the marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 in various scenarios, which are provided for the embodiments of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0032] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0033] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0034] See Figure 1-5 As shown, a marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to an embodiment of this application includes:

[0035] The acquisition module is used to acquire raw image data of underwater marine organisms and process the raw image data to obtain a marine organism dataset.

[0036] The detection module is used to identify the marine organism dataset.

[0037] It should be understood that the process involves collecting raw image data of underwater marine life. The dataset is then augmented and processed to construct a comprehensive marine life dataset with sufficient data and samples. A neural network is used to train the dataset, and the network parameters are adjusted and optimized to obtain the optimal detection model. The trained detection model is then integrated into the detection system to identify marine life. The dataset contains underwater images closely resembling real-world environments. It includes 890 real underwater photographs taken under diverse lighting conditions, covering a wide range of colors and contrasts. Using Python programs, several image augmentation methods were applied to the original dataset, including image scaling, random rotation, image blurring, brightness adjustment, mirror flipping, and noise reduction, expanding the dataset to 5201 images, enriching the data and enhancing the model's generalization ability.

[0038] In some specific embodiments, when the acquisition module processes the raw image data to obtain a marine life dataset, it includes:

[0039] The acquisition module is configured to perform image scaling, random rotation, image blurring, brightness adjustment, mirror flipping, and noise processing on the raw image data.

[0040] In some specific embodiments, when the detection module is used to identify the marine organism dataset, it includes:

[0041] The detection module is configured to identify the marine organism dataset using the VM-Unet model, which includes a PVM layer, a patch embedding layer, an encoder, a decoder, a final projection layer, and skip connections.

[0042] In some specific embodiments, the patch embedding layer divides the input image into non-overlapping blocks and maps them to a high-dimensional space with C channels;

[0043] The encoder consists of four stages, with the number of channels gradually increasing in the order of [C, 2C, 4C, 8C]. Each stage contains a VSS module, which reduces the spatial dimension through the PatchMerging operation.

[0044] The PVM layer divides feature X into four sub-features after normalization by the LayerNorm layer. Each sub-feature is input into Mamba for processing and then optimized through residual connections and adjustment factors. Finally, the sub-features are merged and output. The number of channels for each sub-feature is C / 4.

[0045] It should be understood that the VM-Unet model includes a PVM (ParallelVisionMamba) layer, a patch embedding layer, an encoder, a decoder, a final projection layer, and skip connections. Unlike previous methods, we do not use a symmetrical structure but rather an asymmetrical design. The patch embedding layer is particularly advantageous for underwater image enhancement due to the complexity and variability of underwater scenes. By dividing the input image into smaller, non-overlapping blocks, the model can focus on local features within each block, thereby capturing fine details and unique characteristics that are often distorted or occluded in underwater environments. By mapping these blocks to a higher-dimensional space (default channel count C=96), the model can better represent the complex textures and color variations common in underwater environments.

[0046] Layer normalization helps stabilize the training process by normalizing the embedded blocks, making the model more adaptable to common lighting and contrast variations in underwater images. This normalization step ensures a more consistent and efficient feature extraction process in the encoder. The encoder consists of four stages with progressively increasing channel counts, allowing for multi-scale representation of the image. As the PatchMerging operation reduces spatial dimensions and increases the number of channels, the model can focus on more abstract and larger-scale features at deeper stages. This hierarchical representation is crucial for underwater images because enhancement requires handling both fine details (such as texture) and larger patterns (such as light or shadow). Finally, [2, 2, 2, 2] VSS modules are used in each of the four stages to ensure the model can capture and enhance various aspects of the image at different levels of abstraction. The progressively increasing number of channels in each stage ([C, 2C, 4C, 8C]) allows the model to progressively optimize image quality, highlighting important features, which is crucial for improving the visibility and clarity of underwater images.

[0047] The feature X with C channels is first normalized by a LayerNorm layer and then divided into four sub-features: and Each sub-feature has C / 4 channels. Next, these sub-features are processed by Mamba and optimized through residual connections and adjustment factors to enhance the acquisition of underwater spatial information. Finally, the four sub-features are merged into an output feature Xout through a concatenation operation, maintaining the same C channels. This output feature is then processed by LayerNorm and projection. The specific operations can be represented by the following equation:

[0048]

[0049] Where: Y represents the sub-feature, C is the number of channels, Sp is the split operation, LN represents LayerNorm, Mamba is the Mamba operation, θ is the adjustment factor for the remaining connections, Cat is the concat operation, and Pro is the projection operation.

[0050] The multi-scale feature convolution fusion module, compared to single-scale convolution, can operate on feature maps at different scales, capturing detailed features at different levels in the image, thus significantly enhancing the model's perceptual ability. Introducing both large-scale and small-scale features also effectively reduces overfitting to local details. To fully explore the correlation information between features at different scales in underwater images and avoid simple superposition of features at different scales, we designed a multi-scale convolution fusion module aimed at maximizing the fusion of global spatial information from multi-scale features. This design improves the network's adaptability to complex underwater environments. The specific construction process of this module is as follows: First, image features are extracted using convolutional layers with kernel sizes of 1, 3, 5, and 7, respectively, with each convolutional layer having 16 output channels. The resulting feature maps are then superimposed along the channel dimension and input into the multi-scale convolution fusion module. The input features x1, x2, x3, and x4 are first processed through two convolutional layers, a normalization layer, and an activation function layer, compressing the feature map dimension from C*H*W to C / 16*H*W. Then, the generated feature map is element-wise multiplied with the original input feature map and stacked along the channel dimension. After the multi-scale convolutional fusion module, the spatial size of the feature map remains unchanged, ensuring that the network can extract multi-scale information while maintaining the consistency of global features. Through this fusion module, the network's ability to extract shallow global features of the image is enhanced, thereby improving the model's learning performance.

[0051] In some specific embodiments, the detection module is further configured to extract image features of the marine biological dataset through convolutional layers with kernel sizes of 1, 3, 5, and 7, wherein the number of output channels is 16 in each case. The image features are subjected to two layers of convolution, normalization, and activation processing, and the dimension is compressed to C / 16*H*W before being multiplied element-wise with the original image features and superimposed on the channel dimension.

[0052] In some specific embodiments, the detection module is also configured to identify the marine organism dataset using an optimized YOLOv8 target detection algorithm.

[0053] In some specific embodiments, the detection module is further configured to identify the marine organism dataset using an optimized YOLOv8 object detection algorithm, including:

[0054] A DSC-GAM module is added to the backbone network of YOLOv8. The DSC-GAM module replaces the fully connected layer in the channel attention submodule with depthwise separable convolution to enhance feature extraction and suppress irrelevant information. SCConv is used instead of ordinary convolution in the neck network to reduce computational complexity, and VariFocalLoss is used to optimize the classification loss.

[0055] In some specific embodiments, adding the DSC-GAM module to the YOLOv8 backbone network includes:

[0056] The input feature F1 is processed by a channel attention mechanism to obtain the intermediate state F2. The channel attention mechanism adjusts the channel weights through two MLP layers and depthwise separable convolution.

[0057] The intermediate state F2 is optimized by a spatial attention mechanism to obtain the output feature F3, which highlights key features through max pooling and average pooling.

[0058] In some specific embodiments, the detection module is further configured to employ the AdamW optimizer and a cosine annealing learning rate scheduling strategy, with an initial learning rate of 4×10⁻⁶. -3 Attenuation to 1×10 -6 The training batch size was 16, and the training period was 500 epochs to train the YOLOv8 model.

[0059] It should be understood that the YOLOv8 target detection algorithm was optimized to improve the recognition accuracy of underwater target detection. First, a DSC-GAM module was added to the YOLOv8 backbone network to enhance the model's feature extraction capabilities. Then, SCConv was used to replace ordinary convolutions in the neck network to reduce computational complexity. Finally, VariFocal Loss was employed to optimize the classification loss, effectively addressing the challenge of imbalanced sample classes and focusing on difficult-to-classify positive examples. These innovative technical improvements provide strong support for the effective identification of marine targets. To maximize the effect of the attention mechanism, the DSC-GAM module was implanted between the fourth C3 module and the SPPF module in the backbone network. This configuration allows the improved backbone network to not only utilize the efficient feature selection capabilities of DSCGAM to enhance the network's feature extraction but also effectively suppress irrelevant information, significantly enhancing the detection accuracy of marine targets. Attention mechanisms play a crucial role in improving the performance of computer vision tasks, with compressed excitation attention and convolutional attention being the most common forms. SE focuses on enhancing the importance differences between channels, but its limitation lies in failing to fully utilize spatial information, potentially leading to the neglect of spatial correlations. CBAM integrates channel attention and spatial attention. Both sub-modules utilize max pooling and average pooling to highlight key features of the feature map and consider the overall information of the features. However, this method fails to fully explore the interaction between the channel and spatial dimensions, and max pooling may lead to information compression loss. Although it enhances the representation capabilities of the channel and spatial dimensions, it still has shortcomings in global information integration.

[0060] To overcome these limitations, this patent introduces a global attention mechanism:

[0061] Global Attention Mechanism (GAM) aims to enhance the interaction of global information and reduce the loss of valuable information. This mechanism combines the advantages of CBAM and improves its channel attention and spatial attention submodules, proposing an innovative structure that effectively reduces information loss and enhances feature interaction in the global dimension. The mechanism aims to optimize attention allocation through more detailed global dimension analysis, and its structural design emphasizes enhancing the understanding of global information while preserving local details.

[0062] Through the channel attention mechanism, the input feature F1 is adjusted to obtain target weights, transforming it into an intermediate state F2. Subsequently, the spatial attention mechanism further optimizes these adjusted target weights, and the output feature F3 is obtained through element-wise multiplication. This process enhances the representation of the global target features while preserving key information. The intermediate state F2 and the output feature F3 are shown in the following equations:

[0063]

[0064] Among them, M C and M S These represent channel attention mechanism and spatial attention mechanism, respectively. It performs element-wise multiplication.

[0065] The channel attention mechanism in GAM includes a two-layer MLP, which increases the computational complexity of the module.

[0066] Considering the significant reduction in computational cost of depthwise separable convolution (DSC), DSC was used to replace the second fully connected layer of the MLP in the channel attention submodule. By implementing channel-wise and pointwise convolutions stepwise, the algorithmic complexity of GAM was successfully reduced while maintaining the effectiveness of the attention mechanism. This adjustment makes the GAM module more efficient while reducing the required computational resources, thereby improving the overall network performance while ensuring the algorithm's lightweight and efficient nature. This section names the improved global attention mechanism DSC-GAM.

[0067] To verify the effectiveness of the combined algorithm in this patent, the dataset used in this invention comes from the UIEB dataset, which contains 890 underwater images taken under natural light, artificial light, or a mixture of natural and artificial light, and also provides corresponding high-quality reference images. In this invention, the image data in the dataset is divided into training and test sets in an 8:2 ratio, and the images in the training set are flipped horizontally to expand the training set to 4450 images before training. Finally, underwater images with the following scenes, such as seabed, marine life, and coral, are randomly selected as test images to illustrate the generalization and effectiveness of the model. The original underwater images and their corresponding high-quality comparison images are shown below. Figure 2 As shown.

[0068] Experimental operating environment: CPU is Intel Core i7-14700;

[0069] The GPU is an NVIDIA GeForce RTX 4070 Ti Super, the memory is 32GB, the operating system is Windows 11 Professional (64-bit), and the development software is PyCharm.

[0070] Training parameters: The network input is a standardized preprocessed 256×256 pixel RGB three-channel image. The AdamW optimizer is used for optimization, configured with a cosine annealing learning rate scheduling strategy. Specifically, the initial learning rate is set to 4×10⁻³, and gradually decays to 1×10⁻³ during the training period using the cosine annealing mechanism.-6 The training batch size was kept at 16, and the entire training process lasted for 500 epochs to ensure that the model fully converged.

[0071] (a) Image processing section: To demonstrate the effectiveness of the network model and modules proposed in this invention, two experiments will be conducted. In the comparative experiment, the model proposed in this invention will be compared with the underwater image enhancement methods based on Bayesian theory (Bayesian-Retinex), underwater image enhancement methods based on physical models (ULAP[7], IBLA[8]), Ucolor method, underwater image fusion method, underwater image enhancement methods based on deep learning (UWCNN, CycleGAN), and underwater image enhancement methods that also use multi-scale convolutional fusion technology (Water-Net).

[0072] An objective evaluation was adopted, selecting two full-reference image quality assessment metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). PSNR is the ratio of the energy of the peak signal to the average energy of the noise. PSNR is based on the error between corresponding pixels, i.e., it evaluates image quality based on error sensitivity. A higher PSNR value indicates better image quality. Structural Similarity (SSIM) measures the similarity of images from three aspects: brightness, contrast, and structure, with a value range of [0, 1]. A higher SSIM value indicates less image distortion.

[0073] The experimental results are shown in the figure below. Figure 3 As shown. Bayesian-Retinex and underwater image fusion methods are both local enhancement algorithms, thus suffering from noise amplification during image processing. The CycleGAN algorithm can effectively improve color cast, but the overall color brightness of the enhanced image is darker. IBLA and ULAP methods rely on prior light attenuation, so they do not perform well in restoring severely degraded images. Since the UWCNN method is trained using a synthetic dataset, it cannot fully adapt to real underwater images. The Water-Net method's color correction in strong scattering regions is accompanied by abnormal contrast enhancement (such as saturation overflow of red objects). The Ucolor method produces light spot artifacts in the restored images of waters containing dense plankton. Compared with the above methods, the method proposed in this invention has better restoration results, better robustness, and is effective for both foreground and background of the image.

[0074] In Table 1-4 and Figure 3In the diagram, (a) represents the original image, (b) represents the underwater image fusion method, (c) represents the Bayesian-Retinex method, (d) represents the underwater fog line-based method, (e) represents the IBLA method, (f) represents the ULAP method, (g) represents the UWCNN method, (h) represents the Water-net method, (i) represents the Ucolor method, and (j) represents the method of this invention.

[0075] To verify the above conclusions, this invention introduces reference metrics PSNR and SSIM, as well as non-reference metrics UCIQE and UIQM, to objectively evaluate the processing results of different metrics, thereby further evaluating the quality of the restored images. In Tables 1 to 4, bold text indicates the best evaluation metric, italics indicate the second evaluation metric, and bold italics indicate the third evaluation metric. The meanings of (a)-(j) are consistent with those described above. Tables 1 and 2 show the reference metric evaluation results for the original images and the images processed by the nine methods (b)-(j) from the UIEB dataset. Tables 3 and 4 show the non-reference metric evaluation results.

[0076] In this embodiment, Table 1 is a quantitative comparison table of the reference indicator PSNR: Table 1

[0077] In this embodiment, Table 2 is a quantitative comparison table of the reference index SSIM: Table 2

[0078] In this embodiment, Table 3 is a quantitative comparison table of the non-reference indicator UCIQE: Table 3

[0079] In this embodiment, Table 4 presents a quantitative comparison of the non-reference indicator UIQM:

[0080] Table 4 A comprehensive comparison of the reference and non-reference metrics evaluation results of the images processed by various methods reveals in Tables 1 to 4 that the underwater image restoration method based on second-order statistical properties (j) ranks among the top three in most evaluation metrics. Therefore, the method proposed in this invention performs excellently in terms of the quality of the restored image. Furthermore, since the Water-net method (h) and the Ucolor method (i) use the UIEB dataset as their training set, their reference and non-reference metric evaluation results are also quite good.

[0092] Because in the reference metric evaluation results and non-reference metric evaluation results of the original images and images processed by the nine methods (b) to (j) of the UIEB dataset in Tables 1 to 4, the Water-net method (h) and the Ucolor method (i) use the UIEB dataset as the training set, and they also obtain relatively excellent evaluation results. Therefore, this section needs to further compare the stability of the method proposed in this invention with the Water-net method and the Ucolor method. Figure 4 Line graphs showing the average UIQM and UCIQE for each method on the UCCS dataset are presented.

[0093] like Figure 4 As shown, the horizontal axis represents nine image processing methods, and the vertical axis represents the average values ​​of UIQM and UCIQE. Compared with the other eight methods, the proposed VM-Unet-based multi-scale fusion convolutional underwater image restoration method has relatively higher average values ​​of UIQM and UCIQE. Therefore, the restoration method proposed in this invention achieves better results and higher stability compared with the Water-net and Ucolor methods.

[0094] (b) Image recognition section

[0095] Parameter settings are also crucial for model performance and generalization ability. Appropriate parameter settings can improve model learning efficiency, optimize detection accuracy, reduce the risk of overfitting, and ensure the model exhibits optimal performance under various conditions, thus effectively adapting to complex and ever-changing underwater scenarios. The parameters set in the experiment were: initial learning rate: 0.01; gradient momentum: 0.842; weight decay coefficient: 0.0005; number of warm-up training rounds: 3.

[0096] To comprehensively evaluate the effectiveness of the algorithm, this patent uses the following six metrics as evaluation standards. In the classifier's prediction, TP refers to the number of actual positive samples that are correctly predicted as positive samples. FP represents the number of cases where the classifier mispredicts actual negative samples as positive samples, indicating the number of false positives. Conversely, TN represents the number of negative samples accurately identified by the classifier, while FN represents the number of cases where actual positive samples are predicted as negative samples, reflecting the number of missed positive samples. Precision (P) and Recall (R) are shown in the following formulas.

[0097]

[0098] The F-measure takes into account both precision and recall, and is a harmonic average of the two, used to measure the overall performance of the model. Its formula is shown below. Here, α is a parameter; when the parameter is 1, it becomes the most common F1 score.

[0099]

[0100] Construct a two-dimensional coordinate system where the x-axis represents recall and the y-axis represents precision, and plot a precision-recall (PR) curve based on this system. The area enclosed by the PR curve and the coordinate axes defines the average precision (AP) value. AP is shown in the following formula:

[0101]

[0102] The mean average precision (mAP) is the average precision across multiple classes, where n represents the total number of classes, as shown in the following formula:

[0103]

[0104] Frame rate (Frames Per Second, FPS) is the number of frames per second, a metric that measures the number of images processed per second, as shown in the following formula:

[0105]

[0106] To evaluate the effectiveness of the improvements made to the YOLOv8 algorithm in this invention, this invention trained the original low-quality dataset using the YOLOv8 algorithm before and after the improvements, and tested these trained models. Figure 5The image shows a comparison of the performance of the improved YOLOv8 algorithm in detecting seafood in various underwater environments. The white dashed lines in the image circle areas that the YOLOv8 algorithm failed to detect. The difficulty of detection gradually increases due to color casts, the number of small targets, and occlusion between seafood in different images. In the first row of images, although the color cast is the smallest, the background color is extremely similar to the sea cucumber color, making it easy to miss detection. In the second row, the overall color of the image is light green. In this scene, the shell color of some scallops is similar to the sand and pebbles on the seabed, making them difficult to identify accurately; at the same time, the sea cucumbers and sea urchins located in the center and upper part of the image occlude each other and are also missed by the detection model; on the far left of the image, the sea urchin, which is only half-captured, was also not detected by the model before the improvement. The third row of images has a complex scene due to water absorption and scattering; the crevices between rocks appear dark green in this image, similar to the color of the seafood. Furthermore, the small proportion of seafood in the image due to the long distance shooting further increases the difficulty of detection. The severe color cast and numerous small targets in this scene make accurate detection of sea urchins and starfish difficult. To address these challenges, this invention comprehensively improves the YOLOv8 algorithm, enhancing the model's feature extraction capabilities, optimizing the handling of class imbalance, and particularly improving the ability to identify difficult-to-distinguish positive samples. Figure 5 As shown in the middle and right sections, the improved algorithm effectively solves the problem of missed detection of seafood, accurately distinguishing target objects from complex backgrounds. Seafood that was previously missed is now accurately identified by the improved model. These results fully demonstrate that the improved YOLOv8 algorithm enhances its ability to distinguish between objects with similar background colors and small targets in various scenarios on the original dataset, thereby significantly improving the model's accuracy and robustness in target detection.

[0107] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8, characterized in that, include: The acquisition module is used to acquire raw image data of underwater marine organisms and process the raw image data to obtain a marine organism dataset. The detection module is used to identify the marine organism dataset.

2. The marine biological intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 1, characterized in that, When the acquisition module processes the raw image data to obtain a marine life dataset, it includes: The acquisition module is configured to perform image scaling, random rotation, image blurring, brightness adjustment, mirror flipping, and noise processing on the raw image data.

3. The marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 2, characterized in that, When the detection module is used to identify the marine organism dataset, it includes: The detection module is configured to identify the marine organism dataset using the VM-Unet model, which includes a PVM layer, a patch embedding layer, an encoder, a decoder, a final projection layer, and skip connections.

4. The marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 3, characterized in that, The patch embedding layer divides the input image into non-overlapping blocks and maps them to a high-dimensional space with C channels; The encoder consists of four stages, with the number of channels gradually increasing in the order of [C, 2C, 4C, 8C]. Each stage contains a VSS module, which reduces the spatial dimension through the PatchMerging operation. The PVM layer divides feature X into four sub-features after normalization by the LayerNorm layer. Each sub-feature is input into Mamba for processing and then optimized through residual connections and adjustment factors. Finally, the sub-features are merged and output. The number of channels for each sub-feature is C / 4.

5. The marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 4, characterized in that, The detection module is also configured to extract image features from the marine biological dataset using convolutional layers with kernel sizes of 1, 3, 5, and 7, where the number of output channels is 16. The image features are subjected to two layers of convolution, normalization, and activation processing, and the dimension is compressed to C / 16*H*W before being multiplied element-wise with the original image features and superimposed along the channel dimension.

6. The marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 5, characterized in that, The detection module is also configured to identify the marine organism dataset using an optimized YOLOv8 object detection algorithm.

7. A marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 6, characterized in that, The detection module is also configured to identify the marine organism dataset using an optimized YOLOv8 object detection algorithm, including: A DSC-GAM module is added to the backbone network of YOLOv8. The DSC-GAM module replaces the fully connected layer in the channel attention submodule with depthwise separable convolution to enhance feature extraction and suppress irrelevant information. SCConv is used instead of ordinary convolution in the neck network to reduce computational complexity, and VariFocalLoss is used to optimize the classification loss.

8. The marine organism intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 7, characterized in that, When adding the DSC-GAM module to the YOLOv8 backbone network, the following is included: The input feature F1 is processed by a channel attention mechanism to obtain the intermediate state F2. The channel attention mechanism adjusts the channel weights through two MLP layers and depthwise separable convolution. The intermediate state F2 is optimized by a spatial attention mechanism to obtain the output feature F3, which highlights key features through max pooling and average pooling.

9. A marine biological intelligent detection system based on multi-scale convolutional fusion and YOLOv8 according to claim 8, characterized in that, The detection module is also configured to use the AdamW optimizer and a cosine annealing learning rate scheduling strategy, with an initial learning rate of 4×10. -3 Attenuation to 1×10 -6 The training batch size was 16, and the training period was 500 epochs to train the YOLOv8 model.

Images