Research On Detection Algorithm Of Shallow Sea Treasures Based On Convolutional Neural Network

In recent years,with the increasing demand for holothurian,echinus,shellfish and other sea treasures,the mariculture industry has achieved vigorous development.In the field of marine ranch construction,the methods of pulling nets or manual grabbing are mainly used to fish marine treasures in the underwater environment,but the above methods have problems of high labor intensity,low work efficiency and high safety risks.The underwater robot that can automatically recognizes,locates and grabs marine treasures is of great significance for reducing labor intensity,improving fishing efficiency and promoting the intelligent development of marine ranching.Accurate detection of underwater objects is the basis for automatic detection and grasping of underwater robots.Compared with real terrestrial scenes,the underwater environment is more complex,and limited by special imaging principles,the optical images collected by underwater sensors often have quality degradation problems such as color cast,low contrast,and blurred features,underwater object detection faces huge challenges.Under the above background,this paper studies the underwater object detection algorithm based on convolutional neural network,taking shallow sea treasures as the detection object,a object detection model suitable for the underwater environment is designed and trained.The main work of this paper is as follows:(1)Aiming at the quality degradation problems such as color distortion and low contrast in underwater optical images,a detection model for sea treasures based on SSD and image enhancement algorithm is proposed.In the image preprocessing stage,the multi-scale retinex algorithm with color recovery is used to process the target dataset,so that the enhanced image is closer to the image in the real ground scene,and the generalization ability of the general detection algorithm in the underwater scene is improved;In terms of structure design,a fine-grained detection head is added on the basis of the original SSD structure,and shallow high-resolution feature maps are introduced to form denser predictions,which reduces the missed detection rate of the network;In the network training stage,in view of the large difference in the number of samples between target classes,a class-weighted loss function is proposed to balance the difficulty of training the network for different classes of samples and improve the detection accuracy of difficult objects.Detailed experiments are carried out on the URPU underwater object detection dataset,the test results show that the mean average accuracy of the improved algorithm in this paper is improved by 5.9% compared with the baseline network,and has a higher detection accuracy than the classical object detection algorithm.(2)Aiming at the problems of frame redundancy and unbalanced positive and negative samples in the traditional anchor-based object detection algorithm,the Center Net network based on center point detection is applied to the sea treasure detection task.First,the high-resolution network in the human pose estimation task is used as the feature extraction network,because the internal high-resolution features are always maintained,so the more accurate target position information in the image is retained;Secondly,a parallel bottleneck attention module is introduced after the feature extraction network to enhance the features of the region of interest and suppress useless background or noise information;finally,a feature fusion module is designed to combine multi-scale feature maps to make full use of low-level location information and high-level semantic information inside the network.A detailed comparison experiment is carried out on the URPU underwater dataset,the improved algorithm in this paper has higher detection accuracy and has more advantages in underwater object detection tasks.Additional comparative experiments are conducted on the PASCAL VOC public dataset,compared with the mainstream object detection algorithms,the improved algorithm in this paper also has the advantage of accuracy,which shows the generalization ability of the improved algorithm in this paper.(3)Aiming at the large amount of parameters of traditional object detection network,which is difficult to deploy to mobile and embedded devices,a lightweight sea treasure detection network is designed,and the YOLO v4 network is improved from three aspects: backbone network,neck network and loss function.First,the lightweight Mobie Net v2 network is used as the backbone network,and the depthwise separable convolution is introduced to replace the traditional standard convolution to reduce the amount of network parameters;Secondly,an adaptive spatial feature fusion module is introduced into the neck network,and the fusion weight of each scale feature is adaptively learned,and the useful feature information is retained for combination;Finally,the focal loss is used to replace the standard cross-entropy loss to alleviate the problem of unbalanced positive and negative samples in the network.Trained on the URPU underwater dataset,the algorithm maintains high accuracy,while the model size,parameter quantity and floating-point operation quantity are greatly reduced,which is more advantageous in practical applications.