Research On Object Detection Algorithm Of Marine Biological Based On Deep Learning

Nowadays,underwater vehicles have a wide range of applications and prospects in underwater observation and marine engineering.In order to improve the intelligent level of underwater operation,underwater vision detection technology based on light vision has received more attention.The underwater environment tends to be affected by low light and blur conditions.In such situation,the accuracy and robustness of traditional target detection algorithms have relative bad performances,however,detection algorithms based on deep learning demonstrate its superiorities.Considering the real-time and accuracy demanding of the detection algorithm,this work devotes to realize autonomous detection of underwater natural organisms(i.e.sea cucumbers,seashells,urchins)through using single shot multibox detector(SSD)algorithm,which is a detection algorithm belongs to deep learning.The underwater natural organism data set consists of actual images and augmented images,and these targets are relatively small.Hence the SSD detection algorithm has slightly lower detection accuracy on this data set.Therefore,it is necessary to improve the detection accuracy on the premise of ensuring the detection speed.In addition,images from the target domain have a large difference compared with ones from the source domain,it may cause poor performances when the algorithm only based on the source domain.Moreover,it is often difficult to create a training set by using target images from the source domain.A good model improves the detection performance of environmental changes.In order to solve these problems,the main contents of this work are as follows:(1)The basic knowledge of deep learning is introduced,the SSD algorithm and the Faster R-CNN algorithm are analyzed theoretically.These two algorithms are reproduced in the natural organism data set.Algorithms are optimized by adjusting parameters,and the test results are recorded to judge the results of the proposed algorithm.(2)In order to solve the relatively low accuracy problem of the SSD algorithm in underwater natural organism detection,this work proposes Emphasis on Features-SSD(EFSSD)algorithm.This algorithm introduces a feature emphasis layer in the SSD network,and extracts the color and texture features of the image,and passes them back to the network for joint training to improve the feature extraction capabilities of the algorithm.The proposed algorithm is realized on Caffe,which is one of the deep learning frames.The algorithm is tested and results show that the EF-SSD algorithm not only improves detection effects and ensure the detection speed.(3)To improve the detection performance and solve the problems caused by false detections and omissions in the target source.This work proposes a boosting algorithm of video object detection in target source.Firstly,a video example mining method is introduce based on the high repetitiveness of the video stream images instead of traditional example mining method.Moreover,a hard example data set is generalized through the proposed mining method.Finally,the EF-SSD algorithm model trained in the source domain is automatically retrained,and then the target domain image is re-detected.The proposed method is verified by testing under a period of video which is recorded in the target domain.The results show that the algorithm effectively reduces the false detection and omissions in the video.