Research On Target Detection Based On Improved Convolutional Neural Network

Target detection is the basis of many computer vision tasks,and the research on it plays an extremely important role in the development of intelligent transportation,intelligent monitoring,military guidance,medical surgery and other fields.Most of the early target detection algorithms are based on manually designed features and lack of effective image feature expression methods.In recent years,due to the Convolution Neural Network(CNN)is proposed,effectively overcome the limitations of traditional image target detection method,has been one hot research field of target detection.Based on the careful analysis of the domestic and foreign research status and existing algorithms of target detection technology,combined with the relevant research theory of CNN,this paper improves the current mainstream target detection method based on convolutional neural network to some extent.The main work of this paper is as follows:(1)For the problem that the geometric modeling ability of the convolution unit in the classical CNN is poor,and the ReLU(Rectified Linear Unit)activation function easily leads to the death of neurons,a DP-SSD(Deformable Parametric Single Shot Multibox Detector)target detection method based on adaptive receptive field is proposed.This method in the classic behind each of the pooling of SSD framework layer adds a Deformable Convolution structure,at the same time,a Parametric Rectified Linear Unit(PReLU)is introduced instead of the original ReLU activation function.(2)In order to improve the detection accuracy of Faster R-CNN,a residual network(Resnet)was used to replace VGG-16 as the basic feature extraction network in the Faster r-cnn framework to optimize network performance.At the same time combined with the Depthwise Separable Convolution reduces the network complexity and computation,model checking speeds increase,based on this puts forward an improved Faster R-CNN target detection algorithm.Apply the improved algorithm to the public PASCAL VOC and COCO data sets for target detection experiments.The experimental data analysis shows that the improved algorithm can effectively improve the original algorithm's target miss detection,neuron death,unstable training process,etc.,and improve the detection performance of the algorithm.