Research On Target Detection And Recognition Based On Deep Learning

With the development of computer vision technology,target detection and recognition technology is increasingly applied to people's daily life.The detection and recognition of pedestrians and vehicles is most widely used.But the traditional detection and identification methods are not only difficult to provide high reliability detection result,but also very slow.All these limit the application of target detection and identification technology.Research on deep learning,especially convolutional neural networks,is gradually changing the situation.Therefore,it is of great academic value to study the detection and recognition of pedestrian and vehicle targets and to study solutions based on deep learning.In the target detection task,a very important part is the target windows generation algorithm.Based on the Faster R-CNN theory,I use the intermediate feature map of convolution neural network to generate candidate windows.In order to solve the problem of scale change of the target and to reduce the false detection rate,this paper proposes a multi-RPN layer fusion strategy based on the analysis of the characteristics of the network intermediate convolution feature map.Finally this thesis generates candidate targets on different convolutional feature maps,and enhance the ability to detect targets at all scales.Secondly,aiming at the problem of target position prediction,this paper proposes a logarithmic suppression method and improves the target position regression algorithm,which not only avoids the oscillation problem caused by the excessive error,but also quickens the convergence rate of the objective function near the optimal solution.In order to solve the problem of over-fitting which is prone to convolutional neural network,this paper introduces the methods of avoiding over-fitting in deep neural network,including the introduction of Dropout layer and the effect of regularization term in objective function.Through the distribution of the parameters in the convolution neural network,a method of reducing the dimension of the full connectivity layer is proposed.Then the paper introduces Sparse PCA.And the effectiveness of the algorithm is proved by the experimental comparison data.In the network training phase,the paper analyzes and compares several parameter initialization methods.And their performance is compared and evaluated.The paper also analyzes the principle of gradient descent method.Combined with the experiment,the dynamic process of parameter optimization is clarified,and the adjustment method of coefficient of learning rate is given.At last,the research work is summarized and the next research work is prospected.