Research Of Object Detection Algorithm Based On Improved End-to-End Convolutional Neural Network

The object detection algorithm based on deep convolutional neural network is defined by using deep convolutional neural network to build a model to locate and identify the object of interest in a given picture or video.According to whether the structure of algorithm contains the region proposal module,object detection algorithm based on deep convolutional neural network can be divided into one-stage detection algorithm and two-stage detection algorithm.Both have been widely used in the fields of automatic driving,person re-identification,face recognition,etc.The one-stage detection algorithm has an advantage over the two-stage detection algorithm in speed,but it is relatively inferior in accuracy.By improving the network structure of the algorithm,the shortcomings of the one-stage detection algorithm can be made up.In practice,the object detection algorithm based on deep convolutional neural network has too many parameters and takes up a lot of storage cost.The algorithm can be improved by light-weight methods,which reduce the parameters of algorithm while meeting the requirements of accuracy and speed,thereby improving the problem that the algorithm takes up a lot of storage space.So the research on the object detection algorithms based on deep convolutional neural networks has important value in academic and practical application.This thesis is mainly aimed at improving the structure of the YOLO v3 and FSSD algorithm to make algorithms faster or more accurate.The main work is as follows.1?A improvement of the backbone of YOLO v3 is proposed.The improvement mainly replaces the DarkNet-53 in the YOLO v3 with MobileNet constructed by deep separable to reduce the cost of calculation and improve the speed of algorithm.In addition,the channel attention mechanism is introduced in the MobileNet to improve the ability of network in feature extraction and improve the accuracy of the algorithm.The experimental results on ROSD dataset,LEVIR dataset and NWPUVHR-10 dataset shows that the improved algorithm is faster than the original algorithm and keeps the accuracy of original algorithm.2?A improvement for lightweight FSSD algorithm.This method is mainly aimed at improving the existing lightweight method that MobileNet is used to replace the backbone of FSSD algorithm.This method replaces the backbone of lightweight FSSD algorithm with improved ShuffleNet and replaces the feature fusion method of lightweight FSSD algorithm with BiFPN to make the fused feature has rich semantic information.The experimental results on PASCAL VOC 2007+2012 dataset and Safety Helmet Wearing dataset show that this method improved the accuracy of lightweight algorithm.