Research And Application Of Image Detection Based On Deep Convolution Neural Network

Convolution neural network is an important part of deep learning theory,which has developed from simple several layer network to several hundred-layer deep network.At present,the deep convolution neural network is widely used in the field of computer vision because of its excellent feature extraction ability,especially in target detection using the deep convolution neural network method has become the current mainstream research direction.In this paper,the target detection algorithm based on the deep convolution neural network is studied,and the improved Faster R-CNN algorithm and the improved single shot multibox detector(SSD)algorithm are proposed.The specific work is as follows:1.In view of the poor detection effect of Faster R-CNN algorithm on small targets,occluded or truncated targets and difficult to recognize objects,an improved Faster R-CNN algorithm is proposed.The specific improvements are as follows: Firstly,in the feature extraction module,use Res Net-101 network to replace VGG16 network to improve the feature extraction ability;secondly,in the target location module,more accurate suggestion box generation method and more rigorous regional importance scoring mechanism are used to reduce missed target detection;thirdly,in the target recognition module,the loss value of the hard samples is used for back propagation to improve the recognition effect of the algorithm on the hard samples.The experimental results show that the improved Faster R-CNN algorithm improves the accuracy of target detection.At the same time,the improved model is tested in the automatic driving data set.The experimental results show that the improved Faster R-CNN algorithm has good practicability in the unmanned driving environment perception.2.Aiming at the problems of poor detection speed and large amount of model parameters when SSD algorithm is applied to mobile target detection,an improved SSD algorithm is proposed.The specific improvements are as follows: Firstly,in order to compress the size of the network model,the separable convolutional Mobile Net?Merge BN is used as the feature extraction network;secondly,in order to improve the detection speed without reducing the detection accuracy,the data preprocessing method of scale transformation and random sampling is added;thirdly,in the training stage of the network model,the kernel sparse algorithm is used to reduce the redundant weight and model capacity.The experimental results show that the improved SSD algorithm not only improves the detection speed and reduces the model capacity,but also ensures the detection accuracy,which is suitable for the mobile target detection.3.Aiming at the implementation of the improved SSD algorithm in the mobile terminal,we use Android studio to call Open CV's DNN module to build application software on the mobile terminal,and load the improved SSD network model in the software,so as to realize the real-time detection of the target in the mobile terminal.