Underwater Multi-target Detection Research On Segmentation Algorithm

As for the cognition of the ocean,our country has invested a lot of manpower and material resources to explore and research,but there is still some lack of cognition of the ocean at present,and the competition for underwater resources has become more and more fierce.Therefore,this paper intends to conduct in-depth research on the underwater optical image in multiple targets.Underwater targets are different from clear targets in land remote sensing images.This paper mainly solves the problem of data collection.Secondly,the pre-processing of the collected images is studied,and then the most cutting-edge algorithms are used to identify and segment them.The underwater sea cucumber,sea urchin,scallop and starfish are planned to be detected and segmented.Firstly,based on the characteristics of underwater optical image imaging,data collection and preprocessing are crucial steps,including the current mainstream white balance compensation and dark channel defogging,based on traditional data enhancement methods such as noise enhancement and filtering.Therefore,in the image pre-processing stage,total reflection and gray level method are introduced into the white balance to compare the color compensation effect picture,and then the two fusion operation is carried out to achieve normal color,and then dark channel is introduced to remove fog to increase brightness compensation.Pave the way for subsequent target detection and segmentation.Thirdly,in order to solve the problem of missing and error detection in the detection process of small targets in the seabed,this paper proposes to add a head network,YOLOv5-head4 network,in the YOLOv5 multi-scale prediction.That is a layer of feature extraction is added to the shallow input part of skeleton network to increase the accuracy of small target extraction.In order to tiger weight of YOLOv5,a layer of feature extraction is removed from the deep network by referring to YOLOv4-tiny,that is,YOLOv5-head2 is modified to guarantee the speediness in the training process.Finally,in order to improve the detection intensity of repeated occlusion targets,a flexible non-maximal suppression is proposed to replace the non-maximal suppression.In this way,small target and occluded target can be detected simultaneously.In this paper,YOLOv4++ 2conv5,YOLOv4-tiny and YOLOv4 are compared,which shows that the improved algorithm improves the effectiveness of small target detection and occlusion target detection.Secondly,in order to solve the shortcomings in underwater target detection.We introduced the pixel level segmentation of underwater target and proposed the one-step target segmentation algorithm YOLACT.We replaced the BN layer with GN layer in YOLACT algorithm and introduced weight standardization(WS),which made model training get rid of the dependence on batch training.The convergence speed of the model is faster,the training effect is better,the loss function value is reduced,and the detection performance is improved.Finally,in order to solve the problem of unclear boundary and unclear segmentation effect of target segmentation in segmentation,Mask RCNN instance segmentation algorithmis introduced in this paper.In Mask RCNN algorithm,attention mechanism module is introduced to improve network expression ability,and some training skills and multi-category boundary classification are added.This process not only improves the segmentation index,but also reflects the display of small target overlapping target after example segmentation.It is proved that the improved algorithm improves the accuracy of target segmentation performance.