Underwater target detection method based on improved lightweight YOLOv9

By using the CA attention mechanism in YOLOv9 to optimize the MobileNetV3 network and add a dual-ended attention mechanism, the YOLOv9-CAMobileNetV3DA model is formed, which solves the shortcomings in real-time performance and accuracy of the underwater object detection method, and achieves more efficient marine biological detection .

CN119068322APending Publication Date: 2024-12-03ZHEJIANG SCI-TECH UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411012342.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing underwater object detection methods have shortcomings in real-time performance and accuracy, especially when dealing with marine biological detection, traditional methods are complex in computing and difficult to capture differences between different types of objects, and a single structural feature extractor is difficult to meet complex scenarios. demand.

Method used

The CA attention mechanism is used to optimize the MobileNetV3 network as the backbone network of YOLOv9, and a double-ended attention mechanism is added to its end to form the YOLOv9-CAMobileNetV3DA model to improve the real-time and accuracy of detection.

Benefits of technology

Effectively reduce the amount of model parameters and improve detection accuracy. mAP is improved by 0.63% and 1.39% compared with the lightweight model, and at the same time improves the model's performance in multi-scale and multi-category detection, which is suitable for marine biological detection. Scene.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068322A_ABST
    Figure CN119068322A_ABST
Patent Text Reader

Abstract

The invention discloses an underwater target detection method based on an improved lightweight YOLOv9. The method comprises the steps that data sets of different marine organisms are acquired and preprocessed; a backbone network of the YOLOv9 is replaced with the improved MobileNetV3 network, and lightweight processing is carried out on the improved MobileNetV3 network; adding a proposed double-end attention mechanism at the tail end of the backbone network to enhance feature extraction; training hyper-parameters such as a training round, a batch processing size, a learning rate and an input picture size are set; inputting data of the training set and the verification set into the model for training, and obtaining a group of optimal weight parameters; and inputting test data by using the optimal weight parameter to obtain a predicted visualization result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater target detection, and in particular relates to an underwater target detection method based on improved lightweight YOLOv9. Background Art

[0002] As humans gradually realize the importance of marine resources, more and more scholars are actively exploring and devoting themselves to the research of marine life detection technology. Traditional target detection methods use feature extractors such as SIFT, HOG, SURF, and ORB to extract image features, and then use machine learning methods to classify or regress the features. Due to its high computational cost, lack of real-time performance, poor generalization, and the development of deep learning, traditional methods are gradually being eliminated in the field of marine life detection. In view of the complexity of underwater target detection scenarios and the special needs of processing real-time videos, various feature enhancement extraction methods and parameter pruning methods have been proposed to achieve more efficient and accurate underwater target detectors.

[0003] Guoyan Yu et al. designed a new underwater biological detection network based on YOLOv7, called Underwater-YOLOv7 (U-YOLOv7). This model established a network combining cross-transformation and efficient squeeze excitation modules to enhance the extraction of channel information; before feature fusion, a lightweight CARAFE (Content-Aware ReAssembly of FEatures) operator was used to obtain more semantic information of underwater images; in order to improve the anti-interference ability of the model in underwater recognition, a three-dimensional attention mechanism was also introduced. Zhuoyan Liu et al. proposed a plug-and-play lightweight underwater joint image enhancement module (UnitModule), which was applied to YOLOv5-S to achieve good results. Liu Yang et al. proposed a dual-path (DP) pyramid visual transform (PVT) feature extraction network DP-fishnet. The DP-PVT backbone network consists of a dual-path feature extraction mechanism that integrates Transformer and convolution operations. The visual transformation path is used to capture global features, and the integrated convolutional neural network path is used to enhance the extraction of local features to improve the accuracy of small target detection. Hu et al. used the high resolution of YOLOv4 with more small target information for detection, and replaced the res_unit in the backbone network with dense_unit to detect dense small particles underwater; Zhang et al. added the AFFM attention mechanism to YOLOv4 for feature fusion, obtained richer semantic information, and improved the accuracy of underwater target detection; Li et al. increased the recall rate of small targets by adding corner regression to multi-scale feature maps, and improved the regression loss to reflect the gap between the real box and the predicted box. However, the detector still needs to be further improved in view of the blurred images of this dataset and the similarity between the biological color and the background.

[0004] Although these models have achieved certain success in specific target scenarios, the following problems still exist: the introduced enhanced feature extraction operation unit may achieve lightweight, but the computational complexity of the model backbone is slightly increased or unchanged; on the other hand, the scale of underwater organisms varies greatly, and it is difficult for traditional single structural feature extractors to fully capture the differences between different types of objects, and there is still room for improvement in detection accuracy. Existing underwater detection methods mainly focus on regular square boxes. When underwater organisms change, considering the real-time requirements of underwater target detection, the above existing detection methods are not sufficient to meet practical requirements. Summary of the invention

[0005] The purpose of the present invention is to address the problems of insufficient real-time performance, slow speed and lack of accuracy in existing underwater target detection methods. A marine life detection method is proposed, which uses a MobileNetV3 network optimized with a CA (Coordinate attention) attention mechanism to replace the YOLOv9 backbone network for lightweight processing, and then uses the proposed dual-end attention mechanism to add the YOLOv9-CAMobileNetV3DA at the end of the backbone network.

[0006] The present invention provides an underwater target detection method based on an improved lightweight YOLOv9, comprising the following steps:

[0007] Step 1: Obtain a marine biological target detection data set and perform preprocessing; the marine biological target detection data set contains image data of marine organisms;

[0008] Step 2: Build the YOLOv9-CAMobileNetV3DA model:

[0009] Get the MobileNetV3 network, use the CA attention mechanism to replace the SE attention mechanism in the bneck module in the MobileNetV3 network, and get the improved MobileNetV3 network;

[0010] Get the YOLOv9 network and use the improved MobileNetV3 network to replace the YOLOv9 backbone network to achieve lightweight;

[0011] Add the proposed dual-end attention mechanism to the last layer of the head end of YOLOv9;

[0012] The dual-end attention mechanism includes fusing the output results of the CA attention mechanism and the SE attention mechanism to obtain a YOLOv9-CAMobileNetV3DA model;

[0013] Step 3: Use the marine biological target detection dataset to train the YOLOv9-CAMobileNetV3DA model to obtain the trained YOLOv9-CAMobileNetV3DA model;

[0014] Step 4: Get the image to be tested, input the trained YOLOv9-CAMobileNetV3DA model, and get the prediction result.

[0015] Preferably, the marine life in step 1 includes fish, divers, starfish, corals, sea turtles, sea urchins, sea cucumbers, scallops, squids and jellyfish.

[0016] Preferably, the scale of the MobileNetV3 network used in step 2 is MobileNetV3-Small.

[0017] Preferably, the dual-end attention mechanism in step 2 includes fusing the output results of the CA attention mechanism and the SE attention mechanism through an addition operation.

[0018] Preferably, step 3 also includes the following hyperparameter settings for the YOLOv9-CAMobileNetV3DA model: the training rounds are set to 100, the batch size is set to 16, and the learning rate is set to 0.01.

[0019] Beneficial effects of the present invention:

[0020] Aiming at the special needs of underwater target detection scenes, the present invention takes the YOLOv9 model as a benchmark to make a series of improvement measures, taking advantage of its advantages in real-time, accuracy, multi-scale detection and multi-category detection, and at the same time improves its problems in computational complexity and large number of parameters, and proposes the YOLOv9-CAMobileNetV3DA marine life detection model. The improved MobileNetV3 network is mainly used to replace the backbone network of YOLOv9 for lightweight processing. In order to make up for the problem of reduced accuracy after lightweight operation, a dual-end attention mechanism based on SE and CA attention mechanism is added at the end of the backbone network. The proposed model ensures the accuracy of detection as much as possible on the basis of effectively reducing the number of parameters of the original model. Experiments have proved that the proposed YOLOv9-CAMobileNetV3DA marine life detection model reduces the number of parameters by about 21.1% on the basis of the original model. The replaced CA attention mechanism and the dual-end attention mechanism added at the end of the backbone effectively make up for the loss of accuracy caused by lightweighting, and the mAP is increased by 0.63% and 1.39% respectively compared with the lightweight model. Therefore, the proposed model can effectively cope with the special scenarios of marine organism detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the network structure diagram of the model Yolov9-CAMobileNetV3DA proposed in the present invention;

[0022] Figure 2 This is the network structure diagram of MobileNetV3;

[0023] Figure 3 This is the structure diagram of the bneck module;

[0024] Figure 4 It is the computational flow chart of the CA attention mechanism;

[0025] Figure 5 This is the network structure diagram of the YOLOv9 model;

[0026] Figure 6 It is a calculation flow chart of the dual-end attention mechanism proposed in the present invention;

[0027] Figure 7 It is a line graph of the number of iterations-accuracy of training the model proposed by the present invention;

[0028] Figure 8 It is a line graph of the number of iterations-loss value for training the model proposed in the present invention;

[0029] Fig. 9 It is a set of visualization results predicted by using the optimal weight parameters;

[0030] Fig.10 is the classification confusion matrix of the prediction results of the present invention; DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings.

[0032] like Figure 1 As shown in the figure, an underwater target detection method based on an improved lightweight YOLOv9 refers to a method for detecting marine life by replacing the YOLOv9 backbone network with a MobileNetV3 network optimized with a CA (Coordinate attention) attention mechanism for lightweight processing, and then using the proposed dual-end attention mechanism to add a YOLOv9-CAMobileNetV3DA marine life detection method at the end of the backbone network to enhance feature extraction.

[0033] The input image first passes through many optimized backneck modules (2-12) and the two-end attention mechanism (13) to obtain the initial extracted feature maps of different levels (4, 9, 13). The feature map 13 is then processed by the SPPELAN module and the upsampling operation and the feature Fig. 9 After splicing, feature map 17 is obtained. Feature map 16 is processed by RepNCSPELAN module and upsampling operation and feature Figure 4 The feature map 19 is obtained by splicing. Similarly, feature maps 14, 17, and 19 are downsampled and spliced ​​(20-26) to obtain the final three feature maps of different levels for detecting the target frame and category of marine organisms. In addition, modules 27 to 41 are auxiliary training branches and do not participate in the reasoning detection process. During the training process, the gradient information of the auxiliary branch will be back-propagated to the main branch to help the main branch better learn the characteristics of the target object. This is a major innovation of YOLOv9.

[0034] The present invention comprises the following steps:

[0035] Step 1: Prepare the dataset and perform preprocessing

[0036] The data set used for training the neural network of the present invention is the RUOD data set proposed by Dalian University of Technology. The RUOD data set is a unique underwater target detection data set, which is characterized by covering a variety of general underwater scenes and a variety of underwater detection challenges. The target categories marked in this data set are quite rich, including 10 types of marine life such as fish, divers, starfish, corals, sea turtles, sea urchins, sea cucumbers, scallops, squids and jellyfish. In addition to the conventional training set and test set, the RUOD data set also specifically includes three environmental challenge test sets, which are fog effect, color deviation and light interference. These special test set designs enable the RUOD data set to comprehensively evaluate the performance of the detector under various underwater environmental conditions. This comprehensive evaluation capability makes the RUOD data set have extremely high practical value and research value in the field of underwater target detection. The data in the RUOD data set can better represent the picture situation obtained by sampling actual underwater pictures.

[0037] Step 2: Replace the YOLOv9 backbone network with the improved MobileNetV3 network

[0038] Since the SE module used in MobileNetV3 only considers the information between the encoding channels to measure the importance of each channel, it ignores the position information that is crucial to capturing the structure of the identified object. Therefore, in order to improve the model recognition rate and enhance the model's capture of the location information of marine organisms, coordinate information must be considered. This paper improves the MobileNetV3 model by introducing CA, replacing the SE module in the MobileNetV3 structure with the CA module, thereby optimizing the MobileNetV3 model. The optimized MobileNetV3 network is then replaced with the backbone network of YOLOv9 to achieve lightweight processing of the model. The optimized model has a parameter reduction of about 21.1% compared to the original YOLOv9, making it easier to deploy on embedded platforms with insufficient hardware resources.

[0039] (2.1) MobileNetV3 network

[0040] The launch of MobileNetV1 marks an important innovation in the field of deep learning. Although traditional convolutional neural networks perform well in image processing tasks, they often require a lot of computing resources and memory, which limits their application in mobile devices and embedded systems. To solve this problem, MobileNetV1 creatively introduced the concept of depthwise separable convolution. Subsequently, MobileNetV2 was further optimized on the basis of MobileNetV1. It introduced the inverted residual structure, which can better utilize the advantages of depthwise separable convolution while reducing information loss. In addition, MobileNetV2 also changed the activation function in the structure from ReLU6 to linear activation function to reduce the impact of nonlinear transformation on model performance. MobileNetV3 inherits the advantages of V1 and V2, and introduces innovations such as SE module and h-swish activation function, further improving the performance and efficiency of the model. First, it introduces the SE (Squeeze-and-Excitation) module, which enhances the representation ability of the model by learning the importance of each channel and modeling the dependencies between channels. Secondly, MobileNetV3 updates the swish activation function to the h-swish function, which is a simplified version of the swish function. It retains the advantages of the swish function while reducing the amount of calculation. Finally, MobileNetv3 redesigns the structure of the time-consuming layer, further improving the performance and efficiency of the model by optimizing the bottleneck layer in the network and introducing new efficient layers. The MobileNetv3 model structure is as follows Figure 2 As shown in the figure, the input image first passes through the basic convolution unit composed of convolution, normalization and hswish activation function, and then passes through the stacking process of many core modules bneck, and then performs a global average pooling operation, and finally obtains the output result through two full connection operations. The structure of the core module bneck of MobileNetv3 is as follows Figure 3 As shown in the figure, the original features are increased in dimension by 1×1 convolution, then by depthwise separable convolution, and then processed by the attention mechanism. The improved MobileNetv3 uses the CA attention mechanism to replace the original SE attention mechanism, and finally reduces the dimension by 1×1 convolution. In addition, there is also a residual connection operation that directly connects the original features to the final extracted features.

[0041] (2.2) CA Attention Mechanism

[0042] The CA attention mechanism can not only capture cross-channel information, but also position-sensitive information, giving greater weight to key pixel coordinate positions. At the same time, the computational overhead of the CA module is also very small, which meets the requirements of this article. The CA attention mechanism is different from the traditional two-dimensional global pooling. It splits the channel attention into two one-dimensional encoding processes and integrates the vertical and horizontal spatial information into the attention map. This method can capture long-distance dependencies while retaining precise position information. By multiplying the weighted feature map, the model pays more attention to the target of interest.

[0043] The calculation process of CA is as follows Figure 4 As shown. Assume that the input is a C×H×W feature map, where C is the number of channels, H is the feature map height, and W is the feature map width. CA converts the two-dimensional global pooling operation into a pair of one-dimensional global pooling operations in different directions, and concatenates the feature information in the length and width directions and performs 1×1 convolution to reduce the dimension, obtaining an intermediate feature map of C / r×1×(W+H), where r is the scaling factor used to reduce the number of channels. The intermediate feature map is then decomposed into feature maps in two directions, which are then raised back to the original dimension through 1×1 convolution and the final feature maps in the length and width directions are obtained through the Sigmoid activation function. Finally, these feature maps are combined with the two final directional feature maps through element-wise multiplication to generate the final output feature map.

[0044] Step 3: Add the proposed dual-end attention mechanism at the end of the backbone network

[0045] YOLOv9 adds the dual-end attention module of SE and CA in the last layer of the backbone end to further improve the performance and accuracy of the model, cope with challenges in different scenarios, and integrate the advantages of multiple attention mechanisms. In the YOLOv9 network, since it is in the last layer of the backbone network, it can play an important role in connection. Due to the special bottom-up and top-down structure of the neck end, features of different scales are fully integrated here, laying the foundation for subsequent predictions. Therefore, the structure of the dual-end attention module of SE and CA placed in the last layer of the backbone network greatly affects the performance of the algorithm and effectively compensates for the loss of accuracy caused by lightweighting. The mAP is improved by 1.39% compared with the lightweight model.

[0046] (3.1) YOLOv9 model

[0047] YOLOv9 is a new generation of target detection system jointly proposed by Chien-Yao Wang and others on February 21, 2024. It is the latest product in the YOLO series. YOLOv9 proposed the PGI (Programmable Gradient Information) idea, that is, the problem of gradient information loss in the back propagation process needs to be solved. The network structure of YOLOv9 is as follows Figure 5 As shown, the following are the components of the YOLOv9 model.

[0048] 1. Backbone

[0049] It is responsible for extracting key feature information from the input image and fusing feature maps at different levels to effectively capture the contextual information in the image.

[0050] 2.Head

[0051] Responsible for aggregating the features extracted by Backbone and refining this information through specific strategies, with a special focus on enhancing spatial and semantic information at different scales to improve detection accuracy. And responsible for the final object detection task, including classification and bounding box regression.

[0052] 3. Auxiliary Reversible Branch

[0053] Although the introduction of reversible structure can ensure the integrity of information, it also significantly increases the number of Backbone parameters and increases the computational cost during inference. During the training process, it helps the backbone to better obtain rich return gradient information, making the backbone have higher performance; while discarding this branch during the inference process, the inference process does not increase the time loss. This module is only used in training mode.

[0054] 4. Multi-level Auxiliary Information

[0055] Auxiliary branches are introduced to capture and integrate feature information of different scales and levels, aiming to improve the detection accuracy and performance of the model. At the same time, it also affects the learning process of the main branch through gradient propagation. During the training process, the gradient information of the auxiliary branch will be back-propagated to the main branch to help the main branch better learn the characteristics of the target object. This gradient propagation method helps to alleviate the problem of gradient disappearance in deep networks and improve the training effect of the model. This module is only used in training mode because the gradient it returns is obtained from the auxiliary reversible branch.

[0056] (3.2) Dual-end Attention Mechanism

[0057] The present invention adds a dual-end attention module of SE and CA to the last layer of the backbone end of YOLOv9. The structure is shown in Figure 6 , the reasons are as follows: (1) Improving model performance and accuracy: SE (Squeeze-and-Excitation) and CA (Channel Attention) attention modules play an important role in deep learning models. SE learns the dependencies between channels and adaptively recalibrates the feature responses of channels, thereby improving the representation ability of the model. The CA module can simultaneously consider the attention in the channel dimension and the spatial dimension, allowing the model to focus on more critical areas in the image. The combination of these two attention mechanisms can further improve the performance and accuracy of the model. (2) Coping with different scenarios: SE is suitable for scenarios with a large number of channels, while CA is suitable for scenarios that need to consider attention in the spatial dimension. In the object detection task, the input image may contain a large amount of channel information, and the position and size of the object may also vary greatly. Therefore, the dual-end attention module of SE and CA can better cope with these challenges and improve the robustness of the model in different scenarios. (3) Fusion of the advantages of multiple attention mechanisms: SE and CA are two different attention mechanisms, each with its own unique advantages. By combining them together, we can make full use of their respective advantages and improve the performance of the model.

[0058] Step 4: Set training hyperparameters

[0059] The following table shows the hyperparameter settings for each comparison model. The parameters for training different models are adjusted accordingly based on the scale of the model and the performance of the experimental equipment. The number of training rounds is uniformly set to 100, and the input image size is uniformly set to 640×640.

[0060] Model training parameter settings

[0061]

[0062] Step 5: Input the dataset into the model for training

[0063] According to step 2 and step 3, the YOLOv9-CAMobileNetV3DA marine biological detection model proposed by the present invention is obtained. In step 1, the training data set is processed, and in step 4, the training hyperparameters are set. Next, the training set and the validation set need to be input into the model for training. After 100 rounds of forward derivation, loss value calculation, back propagation, gradient update, and indicator calculation on the validation set, a set of optimal weight parameters is obtained, which is saved for obtaining prediction results.

[0064] Figure 7The relationship between the number of iterations and accuracy is shown. As the number of iterations increases, the model gradually learns the characteristics of the data and performs better and better on the training data, and the accuracy gradually increases. It finally reaches about 0.85. Figure 8 The relationship between the number of iterations and the model loss value is shown. As the number of iterations increases, the model will learn the characteristics of the data better, so the model loss value will gradually decrease, which means that the accuracy of the model prediction results is constantly improving.

[0065] Step 6: Enter test data to obtain prediction results

[0066] Use the optimal set of weight parameters obtained in step 5 to forward deduce new marine life images in the test set, obtain the predicted results and test the final performance of the model. Fig. 9 A set of results predicted by the model are shown. Fig.10 This is the confusion matrix of the model classification. There are ten categories in total, namely holothurian, echinus, scallops, starfish, fish, corals, diver, cuttlefish, turtle, jellyfish and a background category background. The values ​​on the main diagonal represent the correct classification ratio of the model, among which diver, cuttlefish and turtle have the highest correct recognition rates, all higher than 0.9. The values ​​on the non-main diagonal represent the ratio of the model misclassifying one category as another. We can see that the misclassification is mainly recognized as background, and the misclassification between classes is less, and the misclassification is about 0.01. This shows that the possibility of confusing categories in the optimized model is low, and the target is mainly recognized as background.

[0067] The model proposed by the present invention is then compared with other models. The following table shows the comparison of the results before and after lightweighting.

[0068] Yolov9 lightweight test results

[0069]

[0070] (1) Precision: The proportion of correct positive samples among all instances identified as positive samples.

[0071]

[0072] (2) Recall (Recal1): The proportion of all positive samples that are correctly identified.

[0073]

[0074] (3) Mean Average Precision (mAP): a comprehensive measure of the detection performance of the target detection algorithm.

[0075]

[0076] Among them, TP (True Positives) represents the number of positive samples correctly identified as positive samples, FP (False Positives) represents the number of negative samples incorrectly identified as positive samples, FN (False Negatives) represents the number of positive samples incorrectly identified as negative samples, TN (True Negative) represents the number of negative samples correctly identified as negative samples, Q R is the number of categories. After replacing the backbone network with MobiNetV3, the number of parameters of Yolov9 decreased by 12843538, about 21.1%; at the same time, the model accuracy only decreased by about 1.4%, the recall rate decreased by about 3.9%, and the mAP decreased by about 3.3%. It can be seen that while the number of parameters of the Yolov9-MobileNetv3 model is greatly reduced, the model performance has only slightly decreased. The introduction of Mobilev3 shows its effectiveness. While maintaining excellent performance, the number of parameters is significantly reduced, the computational efficiency is improved, and the model is lighter and easier to deploy on various devices.

[0077] The experimental results after optimizing the MobileNetV3 backbone network are shown in the following table. After replacing the SE attention mechanism in MobileNetV3 with the CA attention mechanism, the accuracy is improved to 0.84764, which is about 0.84% ​​higher than the original model; the recall rate is improved to 0.77386, which is about 1.24% higher than the original model; and the mAP is improved to 0.84006, which is about 0.63% higher than the original model. On this basis, after adding the dual-end attention mechanism, the accuracy is further improved to 0.85012, which is about 0.29% higher again; the recall rate is further improved to 0.7833, which is about 1.23% higher; and the mAP is also further improved to 0.85392, which is about 1.39% higher. In summary, the introduction of the CA module and the addition of the dual-end attention mechanism have a positive impact on the target detection performance of the model, thereby improving the detection ability of the model.

[0078] Yolov9-MobileNetV3 optimization experimental results

[0079]

[0080] The following table shows the performance of Yolov9-CAMobileNetV3 combined with different attention mechanisms (SE, CA, ECA and dual-end attention mechanism). Among them, our model is only slightly lower in accuracy than Yolov9-CAMobileNetV3-CA, that is, Yolov9-CAMobileNetV3 combined with CA attention mechanism, but its recall rate and mAP are much higher than Yolov9-CAMobileNetV3-CA. Yolov9-CAMobileNetV3-SE and Yolov9-CAMobileNetV3-ECA are lower in accuracy, recall rate and mAP than our model. This also shows the unique advantages of our model.

[0081] Experimental results of Yolov9-CAMobileNetv3 adding different attention mechanisms

[0082]

[0083] The following table shows the performance comparison experimental results of different target detection models on the RUOD dataset. The target detection models include Yolov8, SSD, Faster R-CNN, Yolov7, Yolov5 and Yolov8-MobileNetv3. Our experiment selected these common and representative target detection models to comprehensively evaluate the performance of the optimized YOLOv9-MobileNetV3 model we proposed. As can be seen from the table, the optimized YOLOv9-CAMobileNetV3DA model performs well in all indicators, especially in terms of precision, recall and average precision (mAP), which exceeds other comparison models. Specifically, the precision of YOLOv9-CAMobileNetV3DA reaches 0.85012, the recall rate is 0.7833, and the mAP is as high as 0.85392, which are all the highest values. This shows that our model has significant advantages in detection accuracy, recall ability and comprehensive performance. In contrast, although the performance of other models is also excellent, they are slightly inferior in some aspects. For example, the mAP of Yolov8 is 0.8208, slightly lower than that of YOLOv9-CAMobileNetV3DA, while the mAPs of Yolov7 and Yolov5 are 0.8074 and 0.80652, respectively, both lower than that of YOLOv9-CAMobileNetV3DA. In addition, although SSD and Faster R-CNN are classic object detection models, their mAPs are 0.5954 and 0.8175, respectively, which are significantly lower than our model. In general, the experimental results show that YOLOv9-CAMobileNetV3DA outperforms other comparison models in various performance indicators, demonstrating its potential and advantages in practical applications.

[0084] Experimental results on Aquarium and RUIE datasets

[0085]

[0086] Aiming at the special needs of underwater target detection scenarios, the present invention takes the YOLOv9 model as a benchmark and makes a series of improvement measures, taking advantage of its advantages in real-time, accuracy, multi-scale detection and multi-category detection, and at the same time improves its problems in computational complexity and large number of parameters, and proposes the YOLOv9-CAMobileNetV3DA marine life detection model. The improved MobileNetV3 network is mainly used to replace the backbone network of YOLOv9 for lightweight processing. In order to make up for the problem of decreased accuracy after the lightweight operation, a two-end attention mechanism based on SE and CA attention mechanisms is added at the end of the backbone network. The proposed model ensures the accuracy of detection as much as possible on the basis of effectively reducing the number of parameters of the original model, and experiments have also proved the effectiveness of the proposed module. In summary, the model is suitable for special scenarios of underwater target detection.

Claims

1. An underwater target detection method based on improved lightweight YOLOv9, characterized in that: The following steps are involved: Step 1: Obtain a marine biological target detection data set and perform preprocessing; the marine biological target detection data set contains image data of marine organisms; Step 2: Build the YOLOv9-CAMobileNetV3DA model: Get the MobileNetV3 network, use the CA attention mechanism to replace the SE attention mechanism in the bneck module in the MobileNetV3 network, and get the improved MobileNetV3 network; Get the YOLOv9 network and use the improved MobileNetV3 network to replace the YOLOv9 backbone network to achieve lightweight; Add the proposed dual-end attention mechanism to the last layer of the head end of YOLOv9; The dual-end attention mechanism includes fusing the output results of the CA attention mechanism and the SE attention mechanism to obtain a YOLOv9-CAMobileNetV3DA model; Step 3: Use the marine biological target detection dataset to train the YOLOv9-CAMobileNetV3DA model to obtain the trained YOLOv9-CAMobileNetV3DA model; Step 4: Get the image to be tested, input the trained YOLOv9-CAMobileNetV3DA model, and get the prediction result.

2. The underwater target detection method based on improved lightweight YOLOv9 as claimed in claim 1, characterized in that: The marine life in step 1 includes fish, divers, starfish, corals, sea turtles, sea urchins, sea cucumbers, scallops, squids and jellyfish.

3. The underwater target detection method based on improved lightweight YOLOv9 as claimed in claim 1, characterized in that: The scale of the MobileNetV3 network used in step 2 is MobileNetV3-Small.

4. The underwater target detection method based on improved lightweight YOLOv9 as claimed in claim 1, characterized in that: In step 2: The dual-end attention mechanism includes fusing the output results of the CA attention mechanism and the SE attention mechanism through an addition operation.

5. The underwater target detection method based on improved lightweight YOLOv9 as claimed in claim 1, characterized in that: The step 3 also includes the following hyperparameter settings for the YOLOv9-CAMobileNetV3DA model: the training rounds are set to 100, the batch size is set to 16, and the learning rate is set to 0.01.

Citation Information

Cited By

  • Target detection method and system based on YOLOv9

    CN120070870A

  • A target detection method and system based on YOLOv9

    CN120070870B