Image Classification Based On Dense SIFT Feature And Its Pooling Model

Abstract:Currently, a large number of digital images exist in people’s lives. These images cover lots important aspects of life. However, lots of the images are unmarked or mislabeled that makes it difficult to be searching, processing and using. Because the traditional image classification method spends a lot of manpower and resources, that computer image classification techniques have been proposed and developed to a high level in a short time. This thesis has summarized existing image classification techniques and proposed a new image classification method. The experimental results show that the proposed method has better classification accuracy than existing algorithms.Convolution neural network is a part of the most popular classification algorithm-deep learning. CNN processes images similar to the neurons, which using Multiple convolution kernel and local connection. So CNN has good applicability. Considering the traditional convolution neural network’s hidden layer does not have any physical meaning and its performance depends on massive training samples, this thesis has proposed a new feature model—dense SIFT pooled model(SP model) and constructed dense SIFT pooled convolution neural network (SPCNN) with the model. The classifier uses a Gaussian kernel function instead of the traditional convolution filters to extract the dense SIFT features from images, and gets the SP model which described object characteristic and distribution through the inner product space pooled. This thesis improves SPCNN classification accuracy with a SP model is superior to the traditional classifier by experiments.In order to get better classification accuracy, this thesis uses SP model of the image. So this thesis has proposed a classification method named dense SIFT pooled model based random forest (SPRF). For a better determining the local objects, SPRF algorithm includes the sampling and filtering the image space, making it to be new subspaces. Compared with a normal random forest classifier, SPRF uses stronger base classifier, and gets a smaller generalization error. Experimental results show that SPRF algorithm which has used the SP model has better accuracy than general SIFT features, general random forest and integration algorithm classification.Contributions of this thesis is constructing SP model using dense SIFT features and proposing SPCNN and SPRF classification algorithm based on the SP model which have enhanced the accuracy of the existing image classifier. It will have a certain reference value in image classification.