The Research On Image Target Recognition Based On Deep Learning

Target recognition is an important field of computer vision.With the rapid development of China's urbanization,smart city and the intelligent transportation system has become the trend of development,which puts forward urgent needs for the rapidity and accuracy of target recognition.With its strong feature learning ability,deep learning has achieved remarkable results in many fields of computer vision.Therefore,the target recognition algorithm based on deep learning model has become a research hotspot in recent years.In this paper,a series of researches are carried out on how to identify and locate the image target using the method of deep learning.(1)The common target recognition algorithm is compared and the principle of various algorithms is described.Their advantages and disadvantages are analyzed.This paper summed up the relationship and difference between the selection of candidate region,feature extraction and feature classification.(2)The principle of YOLO algorithm is studied.Based on the tiny-yolo model,an improved model m-yolo was designed.The m-yolo model improves the feature extraction capability by increasing the number of convolutional layers.At the same time,in order to ensure that the improved model will not decrease in the recognition speed,the NIN(network in network)convolution layer is added in the model to reduce the parameters of the model and the complexity of the model.And the function of the NIN convolution layer in improving model is described.Experiments show that the improved model improves the accuracy of recognition and localization.(3)The migration learning strategy is used to improve the performance of the model.The model of random initialization network parameters is compared with the model which uses the pre-training model to initialize network parameters.The experimental results show that the model which uses the pre-training model to initialize network parameters improves the accuracy of recognition and localization.(4)The performance of ReLU and Leaky function was studied.Experiments show that the performance of Leaky function in the vehicle identification task described in this paper is better than the ReLU function.(5)The effect of Anchor Box on model performance was studied.Use the unsupervised learning algorithm k-means to cluster the boundary boxes in the vehicle data set and obtain a Anchor Box with appropriate size and shape.Train the models with different Anchor Boxes.Finally,the optimal number of Anchor Boxes is determined by experiments.