Research On Learning Methods For Identifying Typical Underwater Target Images

UUV can replace traditional submersible in many complex waters to perform monitoring,detection,and other tasks,which has been applied into many military and civilian fields.Image information is used to identify targets,so image recognition technology is essential for UUV autonomous operation.Most of the current image recognition technologies are composed of manual image features extraction,classifier designing and feature recognition.However,the deep learning method subverts the traditional way of image recognition.Deep learning network automatically extracts image features and recognizes them without human participation.The underwater vision image recognition system based on deep learning is mainly composed by these parts: underwater image acquisition,image learning library establishment,underwater image pre-processing and underwater image recognition.In this paper,the establishment of underwater image learning library,the enhancement and denoise propose of underwater image and underwater image recognition are studied.Firstly,the characteristics of underwater typical targets are analyzed,and geometric objects are used instead of real targets for image recognition.Aiming at the lack of visual images of underwater typical targets,A typical underwater target image generation model based on generation adversarial networks is constructed.New images are generated by using a small number of original images,extend image learning library effectively.Secondly,aiming at the characteristics of underwater image data such as large volume,sensitivity to noise and poor contrast,the method of image grayscale,median filtering and histogram equalization are studied.Experiments show that these methods can effectively reduce the amount of image data,reduce the impact of noise and increase the contrast of the image,thus reducing the difficulty of the next experiment.Thirdly,a deep neural network is constructed to identify underwater typical target images.Processed image learning library is used as data source,the number of neurons in hidden layer is modified,Selu function is used as activation function,Adam method is used as learning method,and the network is built by Tensorflow.The final test results show that model recognition accuracy is high.Finally,the effects of different structures,convolution kernels and pooling methods of convolution neural network on underwater typical target image recognition are analyzed.Then use the same activation function and learning method in the previous chapter,a four level convolution neural network model with 5*5 convolution kernel and maximum poolingmethod is built.The final test results show that the recognition accuracy of the model is very high,and the test achieves the desired effect.