Image Classification Algorithm By Combining Mid-Level Features Via Multiple Kernel Learning

Image classification is an important research problem in the field of computer vision.It is widely used in image analysis tasks such as natural scene understanding and industrial detection.The study of high performance image classification algorithm is meaningful for large-scale image management and effective organization.The image feature representation directly affects the performance of the classification algorithm.However,in the classification algorithm based on the local features,the relationship between the local features is ignored,and some information may be lost when a single feature is used to represent the image.At the same time,with the popular of convolutional neural network,it is necessary to fuse traditional image features and convolution features to improve the classification accuracy of images.To address the above problem,this paper has done two studies:1.Construct the mid-level features based on local features.Specifically,the mid-level features are extracted effectively from the local features in the superpixel region,and the structural information of local features is embedded.Then local features and mid-level features are combined under the sparse coding framework to make features have better discriminating ability.Compared with the representative image classification algorithms which only use local features,the new method improves the classification performance of about 3-4% on several common benchmarks.2.In order to further improve the expression ability of local features and mid-level features,kernel functions are used for local features and mid-level features respectively,and the corresponding coefficients are learned to further improve the performance of image classification.Under the multiple kernel learning framework,the local,mid-level and convolutional features are fused.It is verify that these features are complementary on the image classification task.The experimental results show that the accuracy of image classification is improved by 4-6% after the feature of convolutional neural network is added.