Application Of Convolution Neural Network In Target Contour Recognition

Contour detection is to extract the contour of the target object in the image,ignoring the complexity of the background,a variety of textures and noise effects.It is the basic components of the target detection,tracking,recognition of the composition.The traditional contour detection methods are all based on edge detection,which can detect the edge by identifying the brightness point that changes significantly in the image.However,this low level edge detection is carried out in an ideal situation,it is difficult to get a complete target contour in practice.In this paper,It can directly process the pixels of the image to detect the target contour by using convolution neural network,while the extracted contour can be more refined by using the high-level features.First of all,considering the use of classical deep learning framework,feature extraction and classifier combination for contour extraction,the classification error between contour classes can be tolerated,however,the classification error between contour classes and background classes can not be tolerated.Thus,the Softmax loss function is improved and the target profile is obtained.However,the result of detection is general.Then,it can be seen that multi-scale extracted contour will be more refined,by introducing the idea of multi-scale and comparing single-scale and multi-scale network structure to analyze the advantages of multi-scale convolution neural network.Finally,by using VGG,a classic convolutional neural network based on multi-scale,and improving it.Furthermore,by abandoning the full connection to the network structure,the entire reel is used to predict the profile of the target.We can get very good results.And this article,for the small number of data sets,training prone to over-fitting problem,this paper gives a certain solution.The article was conducted on the BSDS500,a data set compiled by the University of Berkeley.At the same time,this article also experiments on its own dataset.Considering the accuracy and recall rate of the experimental results to measure,the algorithm has some advantages over other methods.