Underwater Object Detection And Segmentation Algorithm Based On Deep Learning

As the population continues to grow,terrestrial resources are being consumed at a faster and faster rate.Organisms originate from the ocean,and the ocean is also the most resourceful biological treasure trove on earth.In order to better exploit the ocean resources,human beings have gradually started to use underwater robots to explore the ocean in a series of ways.Underwater semantic segmentation is one of the key technologies of underwater robot vision system,and its main task is to segment the underwater object at pixel level to get the specific outline of the underwater object,so that the underwater robot can work underwater more safely.Due to the complex underwater environment with poor lighting,small objects and severe occlusion,most of the semantic segmentation algorithms at this stage are only used in the above-water environment,while the development of underwater semantic segmentation algorithms has been relatively difficult.To address the above problems,this paper proposes an underwater semantic segmentation algorithm that realizes multi-object segmentation by multi-branch single-object segmentation based on the experimental results that single-object segmentation accuracy is better than multiobject segmentation accuracy,and combines two algorithms,object detection algorithm and semantic segmentation algorithm.The main contents of this paper are as follows:(1)For the problem of more small objects underwater,this thesis proposes an improved YOLOv5 underwater object detection algorithm by using the YOLOv5 network,which is relatively well established at this stage,as the basis of the algorithm.The improved CBAM attention mechanism is added to the feature extraction layer to improve the feature extraction ability of YOLOv5 network;the feature fusion layer is improved to increase the amount of information contained in the fused feature map,and the network loss function and learning rate adjustment mode are improved to better accelerate the convergence speed of the network.Experiments show that the m AP value of the improved YOLOv5 algorithm in this thesis can reach 83.9%,which can detect underwater objects,especially small underwater objects,more accurately.(2)To address the current situation that there are few underwater semantic segmentation algorithms,this thesis proposes an improved Deep Lab v3+ underwater semantic segmentation algorithm using Deep Lab v3+ network,which is a better semantic segmentation algorithm at this stage,as the basis of the algorithm.A feature extraction network suitable for Deep Lab v3+ network is selected among Mobilenetv2 network and Xception network,and hybrid cavity convolution is utilized in the cavity space convolution pooling pyramid structure instead of ordinary cavity convolution calculation.It is proved that Mobilenetv2 network as the feature extraction layer network segmentation of Deep Lab v3+ network is more effective,and the hybrid cavity convolution can increase the amount of information contained in the feature map while keeping the perceptual field unchanged,and then improve the accuracy of underwater image segmentation.(3)To improve the accuracy of underwater multi-object segmentation,this thesis first proposes a YFD(YOLOv5 and FCN-Densenet)underwater multi-object semantic segmentation algorithm,and at the same time,based on this algorithm,combines the improved YOLOv5 underwater object detection algorithm and the improved Deep Lab v3+ underwater semantic segmentation algorithm,and proposes a semantic segmentation using multi-branch single object segmentation The semantic segmentation algorithm IYID(Improved YOLOv5 algorithm and Improved Deep Lab v3+ algorithm),which realizes multi-object segmentation,is proposed.The algorithm firstly inputs the images into the improved YOLOv5 underwater object detection algorithm,secondly saves the confidence frames from the object detection results as confidence frame images and classifies them according to categories,then inputs all confidence frame images of each category into the improved Deep Lab v3+ underwater semantic segmentation network that segments only that category,and finally fuses the segmentation results of all categories to obtain the multi The results of the multi-object segmentation of this image are obtained by fusing the segmentation results of all categories.The experimental results show that the m Io U value of this algorithm can reach 68.80%,which is higher than the accuracy of general semantic segmentation algorithms in underwater segmentation,and also proves the effectiveness of this algorithm.