Research On Underwater Target Recognition Based On Convolutional Neural Network

Underwater target recognition is widely used in military and civil fields such as seafloor salvage,resource exploration and exploitation,and pipeline laying.Optical vision makes up for the deficiency of acoustic vision due to its advantages of close-range perception and richer image information.Therefore,it has become one of the most important branches of underwater target recognition research for domestic and for eign scholars in recent years.This thesis focuses on the in-depth study of recognition technology based on optical vision.And the main work is as follows:(1)The hardware platform of the underwater target recognition system is desig ned and set up,and the data sets are made and labeled.Two types of samples,whi ch are work pieces and geometric objects,are selected as target samples,and the dataare collected by using the established experimental system.A data enhancement method based on image style transformation is provided to expand the data set.All images of the data set are labeled with labeling tools to generate a label file in PAS CAL VOC format.(2)On the basis of analyzing two kinds of commonly used image enhancement methods,a new image enhancement method based on improved global background estimation and color correction is proposed to eliminate image degradation and color decay.During the global background light of an image is estimated,a rectangular template is employed to split the image into different blocks to calculate the color sat uration variance,and the block with the least variance is regarded as the estimated background.Because the original background light estimation method will whiten theimage,the processed image is filtered with the minimum filter.Color of the R cha nnel of the image is corrected by the Retinex algorithm,and the color attenuation coefficients of each color channels are combined.Finally,comparison experiments are carried out to verify the superiority of the proposed algorithm.(3)According to deep analyzation of the common feature extraction methods named as SIFT and HOG,an improved Faster R-CNN is provided to overcome the shortcomings of traditional target recognition methods,such as large computational load,the need for manual feature extraction.The feature extraction layer of Faster R-CNN is reconstructed by using deep separable convolution,and by replacing convolution layer with the full connection layer.Hence,the number of parameters is evidently reduced in the network.ROI Align is adopted to reduce the loss of accuracy caused by the two-quantification of Faster R-CNN,which improves the accuracy of target recognition consequently.(4)Experiments are carried out,and the proposed methods are also evaluated qualitatively.The results indicate the effectiveness of the work done in the thesis.