Research On Detection Method Of Underwater Small Biological Target Optical Image Based On Deep Learning

Marine living(aquatic) resources are important sources of human food.Common aquatic products include scallops,sea urchins,sea cucumbers,etc.,which have high medicinal and nutritional value.The common fishing method is artificial fishing,which has low fishing efficiency and great harm to the fishermen.Therefore,intelligent fishing with underwater robots has become a development trend.Target detection algorithm is one of the key technologies of intelligent fishing.At present,the mainstream target detection algorithm based on deep learning is to detect common targets in life,which can not directly detect small underwater organisms.Therefore,this paper studies the target detection algorithm based on deep learning,designs and implements the underwater small biological target detection algorithm based on deep learning.The small-scale underwater biological data set established in this paper takes the data set of 2018 underwater robot target grabbing Competition preliminary competition as the primary data set,carries out data cleaning and data annotation,and divides the data set into training set,verification set and test set according to 6:2:2.The targets to be tested are scallops,sea urchins and sea cucumbers.The statistical characteristics of the data set are analyzed.The data set in this paper has the characteristics of complex image,rich content,small target to be detected and occlusion.Aiming at the problems of color deviation and atomization in underwater image,the gray world algorithm is used to denoise,and the dark channel prior algorithm is used to defog.This paper studies the fast RCNN algorithm in the region based deep learning target detection algorithm,and improves it according to the characteristics of the data set in this paper.In the selection of the backbone network,resnet101,which has strong feature extraction ability,is selected,and the feature maps of the three convolution modules after resnet101 are fused.At the same time,in order to enhance the real-time performance of the region based algorithm,the backbone network is tried to be replaced For mobilenet-v1,the initial anchor generation in RPN network is changed from the initial value of the algorithm itself to that generated by clustering algorithm,and the ROI pooling operation in fast RCNN is changed to ROI align.Yolo-v3 and SSD in the algorithm of deep learning target detection based on regression are studied and improved according to the characteristics of the data set in this paper.For SSDalgorithm,the backbone network is replaced by resnet101,and the NMS in post-processing phase is replaced by soft NMS.For yolo-v3 algorithm,the focal loss is referenced in the loss function,and the NMS in the post-processing phase is replaced by soft NMS.Through the contrast experiment,it is proved that the improved operation of this paper can improve the detection accuracy,detection frame accuracy,or detection speed.The fast RCNN improved algorithm based on resnet101 as the backbone network has the best performance in terms of detection accuracy and detection frame accuracy,but the real-time performance is poor.The improved SSD algorithm has the best real-time performance,while maintaining high detection accuracy and detection frame accuracy.