Feature Selection Method For Image Classification

Image classification is an important part of computer vision tasks,and the key is how to make feature selection.It is the transformation from traditional manual feature extraction to convolutional neural network(CNNs)depth feature extraction that makes the image classification result improved significantly.Most of the existing CNNs networks use the whole image feature as the basis of image classification,and seldom consider spatial information of image.Spatial information of image is very important for classification task.Therefore,we consider spatial information of applied image to improve classification result,and divides spatial information of image into saliency information and background information.In this thesis,a multi-objective weakly supervised localization algorithm is proposed,and spatial information of image is extracted by weakly supervised localization algorithm.After obtaining spatial information of image,we further propose a classification algorithm based on image spatial information fusion.The main work of this paper is as follows:(1)A multi-objective weakly supervised localization algorithm BCAM(Binary Class Activation Mapping)is proposed.Aiming at the problem of inaccurate localization of CAM(Class Activation Mapping)algorithm on multi-target dataset,a BCAM algorithm is proposed.The BCAM algorithm improves target mapping unit and loss function of CAM algorithm.On the target mapping unit,BCAM sets up separate convolution mapping units for each category,with each category as a two-category problem.In the loss function,in order to calculate the loss for each target mapping unit,BCE(Binary Cross Entropy)loss function is used to optimize network parameters.Through comparative experiments,it is verified that the location result of BCAM on multi-target data sets is more accurate than that of CAM algorithm.(2)An image spatial information based feature selection method is proposed.We select the feature based on the heat map generated by the weak supervised positioning algorithm to extract the spatial information of the image,and further acquire the features of the significant region and the background region according to the heat map,and process and fuse the features of different regions.In feature processing and fusion,a variety of methods are used to conduct experiments.Finally,the fusion features are used to train SVM classifier.The comparative experiments with existing classification algorithms demonstrate the effectiveness of our feature selection method.