Image Classification Algorithms With/without Deep Learning

As an important branch of artificial intelligence,computer vision has found its applications in practice,such as face recognition and target tracking.A basic problem in computer vision is image classification,and it is of great importance to implement image classification efficiently and accurately.Therefore,this paper mainly focuses on image classification algorithms,which involve unsupervised image classification and supervised image classification.Unsupervised image classification does not need labeled data,but calculates the similarity between images according to their features,and then completes the classification.Supervised image classification,however,requires a large number of labeled training data to train model,and then use the trained model to predict the categories of unknown images.For the unsupervised image classification,we propose two methods,which are all based on clustering algorithms.One is the unsupervised image classification by incorporating SIFT(Scale-Invariant Feature Transform)features into AP(Affnity Propagation)clustering algorithm,and the other is based on the hierarchical clustering algorithm.The contribution on the unsupervised image classification can be summarized as follows.(1)With the keypoint matching procedure in SIFT algorithm,we propose a soft matching method to replace the original hard matching method,and the experimental results on ORL face database verify the validity of this new method.Whether with AP clustering algorithm or hierarchical clustering algorithm,the classification accuracy of using soft matching is obviously better than that of using hard matching method.(2)New metric of similarity between images after SIFT keypoint matching is proposed.In the unsupervised image classification based on the hierarchical clustering,we put forward two kinds of new method to define the similarity between images,and the experimental results show that the proposed methods have significantly improved the classification accuracy.For the supervised image classification with deep learning model,we study the famous convolutional neural network(CNN)and the multi-grained cascade forest model based on decision tree(gcForest).Through experimental comparison and theoretical analysis,the two deep learning models are compared comprehensively.Experimental results show that the gcForest model has fewer parameters than CNN,and it does not require a heavy parameter tuning process to get the similar classification accuracy to the optimal CNN model.In addition,the model does not have to use a lot of training data like CNN,and gcForest can also get better classification results on small-scale datasets.