Research On Image Detection Technology Based On Convolution Neural Network

At the end of nineteenth Century,the invention of car continuously development of human society.Today,as far as the application of automatic control technology in the car,relevant technology of deep learning research has been breakthrough.Autonomous vehicles become a major change in car employ related technologies.Deep learning research and application in computer vision is very extensive,such as security and technology for autonomous vehicles,which can perceive the surrounding environment and things change is urgent problems of computer vision.Comparative analysis of the algorithm and image detection based traditional image detection algorithm for convolutional neural network in two aspects.On the basis of explaining the basic theories of traditional image processing,this paper introduces gradient direction histogram HOG algorithm and scale invariant feature transform SIFT operator to carry out principle analysis and experimental analysis and demonstration.The traditional object detection based on the framework is composed of gradient direction histogram and feature classifier.This paper will demonstrate through experimental analysis,correlation algorithm,detection algorithm of some object in the image is not able to detect and mistake similar object detection,affect the detection accuracy and effect of the algorithm.Therefore,this paper proposes a convolutional neural network based image detection algorithm,which greatly reduces the defects of early classical algorithm in image and improves accuracy and recall rate.In order to improve the accuracy of detection algorithm,we are put forward four improvements based on target detection algorithm based on Faster-RCNN,including image augmentation algorithm.This paper proposes a random erase algorithm helps to enhance the robustness of the model and the test sample performance.based on improved feature extraction algorithm on the original network using the feature extraction,the deeper layer network,from the original 16 layers expanded to 101 layers of residual layer,further help convolutional network model to extract the semantic information of the image in the deeper,improve the algorithm in image detection process,the accuracy of semantic information classification and object bounding box.Using different size images the input to the algorithm model,convolutional network feature extraction of Multi-Size-Image after fusion,the algorithm of the model to learn semantic features when the same category can maintain similarity and maintain the maximum difference among different classes of image semantic features;optimize the algorithm area candidate Box feature extraction module,effectively predict the probability of the border categories and target classes of target categories,and improve the accuracy of target detection.In addition,in order to improve the algorithm of model test time,based on the improved Faster-RCNN target detection algorithm,the algorithm by redesigning the ROI pool operation and eliminate the target detection algorithm in the connection layer,so that the model has the invariance properties of target translation,and effectively improve the speed of image target detection model algorithm.At the end of this paper,according to the description of the image detection algorithm of the improved image highlights detection field advantage in application.This article set up the CAFFE and TENSORFLOW deep learning image detection experiment platform based on the framework of the traffic sign by the German Institute of neural computing detection benchmark public data set as the training set and test the model set.The final results can be used to know the image algorithm greatly improves the accuracy and speed of model and fast image detection algorithm is proposed in this paper can improve the model.