Research On Target Detection Method Of Dolphins Based On Deep Learning

In recent years,attacks and information theft in the marine field have been gradually completed by bionic robots.Because the appearance color of dolphins is very close to that of sea water,it has a very high concealment.Therefore,bionic robots often appear as bionic dolphins.The research on dolphin target detection is helpful to understand the bionic characteristics and movement characteristics of the bionic dolphin,and provides a basis for the research of bionic dolphin target detection methods,and is of great significance to the improvement of my country’s underwater defense system.At present,there have been extensive and in-depth research on target detection algorithms based on deep learning at home and abroad,but there is still a problem of lack of training data in target detection tasks for dolphins.At the same time,the collected dolphin images will be severely affected by factors such as illumination and noise in the complex underwater environment,and it is difficult for the target detection model to perform feature extraction on such images.Therefore,this thesis studies the dolphin target detection from two perspectives:data augmentation and enhancement of the feature extraction ability of the target detection model.The main work of the thesis is as follows:(1)Aiming at the problem of lack of training data for dolphin target detection task and poor discrimination of underwater optical image front and background,which leads to the difficulty of edge recognition,a mixed data augmentation scheme of dolphin image based on front and background separation is proposed.This solution firstly binarizes the image to obtain the contour of the object to be detected,and then extracts the object to be detected from the original image and saves it as a new image.The new image and the original image are combined to form a new data set.Experiments prove that the dolphin image hybrid data augmentation scheme proposed in this thesis can effectively alleviate the lack of data.The model trained with the new data set has a significant improvement in the comprehensive performance of detecting dolphin optical samples,which proves that the scheme is highly effective Sexuality and practicality.(2)Aiming at the difficulty in extracting features of dolphin images,a jump route multiscale multi-sense feature fusion model based on YOLOv3(Jump Route Multi-scale Multi-sense YOLOv3,JMM-Y3)is proposed.First,by improving the residual block of the convolutional neural network in the model,the model feature extraction ability is improved;secondly,the target dynamic perception module is proposed based on deformable convolution,which improves the model’s adaptability to non-static objects;finally,the use of regularization The optimization method prevents the model from overfitting during the training process.Experiments show that the JMM-Y3 model has a significant improvement in recognition accuracy and recall rate compared to YOLOv3.At the same time,after applying the data augmentation scheme proposed in this thesis on the JMM-Y3 model,the performance of the model has been significantly improved.