Research On Underwater Environment Image Recognition Based On Transfer Learning

With the advancement of China’s marine strategy,the status of marine economic development and scientific research is increasing day by day.Marine equipment is developing rapidly in the direction of automation and intelligence.Among them,underwater visual equipment,as an intuitive means for human beings to understand the underwater world,has been paid more and more attention.Aiming at the cumbersome process of manually designing feature extractor in traditional machine learning and the problem that depth network can not obtain efficient performance on small sample data sets,this paper studies an intelligent underwater image recognition method based on five kinds of fish image data sets and convolution network.The main work of this paper is as follows.Firstly,in order to solve the problem of insufficient number of fish data sets in the target domain,on the basis of collecting static pictures,this paper intercepts dynamic video frames to obtain static pictures,and uses data enhancement methods to expand the data scale.Due to the low contrast caused by the dark color of the underwater environment image,the image enhancement method is used to improve the image contrast.Aiming at the problem of noise in underwater image caused by impurities in water,a median mean joint filtering method based on set consistency decomposition is proposed to reduce image noise.Secondly,aiming at the problem that it is difficult to obtain a good network model by training convolutional neural network under the condition of limited data set,this paper introduces the direct transfer learning method to improve the recognition rate of network underwater image.Vgg 16 and inception_The simulation results show that the convolutional neural network obtained by direct transfer learning improves the recognition rate of underwater image and reduces the computational cost compared with the network obtained by conventional training.At the same time,in order to reduce the complexity of network model parameters,the full connection layer of convolutional neural network is refined and simplified.The simulation results show that the accuracy of the trained model is more than 90%,which can meet the practical application requirements of image recognition.Finally,in order to solve the problems of data balance and negative transfer in direct transfer learning of small sample underwater image data set,an image recognition method based on secondary transfer learning is proposed on the basis of direct transfer learning.Cifar-10 is used as an intermediate transition data set for secondary transfer learning,which improves the accuracy of underwater image recognition network model.Aiming at the problems of many parameters,slow training and large storage of secondary migration network,the full connection layer of convolutional neural network is also refined and simplified.The simulation results show that the accuracy of the simplified model is 3.33%higher than that of direct transfer learning.