Image Classification And Retrieval Based On Visual Features And Machine Learning

In recent years, with the rapid development of both Internet and multimedia technology, image classification and image retrieval have become worldwide research hotspots. The traditional image retrieval based on key words can not meet the requirements. In that case, a new image retrieval based on content comes into being. There exist mainly two representative methods for image feature description, namely the image visual feature and abstract semantic feature extracted by machine learning. However, the conventional visual feature and abstract semantic feature can not achieve satisfactory results in classifying and retrieving images. To overcome this problem, an algorithm for image classification and retrieval based on visual features and images features extracted by convolution neural network(CNN) is proposed in this paper.Firstly, the paper design a post-processing frame based on bags of words model. Taking advantages of CNN features,the frame calculates the similarity of the images, and then imports the images with high similarity into classifier. These imported images are trained by BOW clustering model. The image retrieval result is regarded as the category which has the largest image number. Under the same data set and the same classification types(this paper applies six types of images whose total number is 1000), experimental results show that the designed frame can obtain the image retrieval accuracy of 90.4%, which improved by ten percent compared with the retrieval algorithm based on CNN features.Secondly, the paper presents an image retrieval method based on color feature and machine learning in detail. Specifically, the algorithm extracts the color features and CNN features separately. The fusion of two weighted feature vectors is treated as the images' descriptors. The similarities of images are calculated through the fused feature vectors so as to acquire the final image retrieval results. It can be shown from the experimental results that the algorithm can improve the image matching degree in color between the retrieval results and imported images dramatically.Finally, an algorithm based on multiple kernel learning and image visual features is presented. Taking use of CNN features, the local self-similarity features and the tower keyword histogram features, the algorithm calculates the corresponding precompiled kernel matrix respectively. The next step is solving the synthesis kernel function and the weights of basis kernel function to get the image classifier used to image classification. Experimental results show that the algorithm based on multiple kernel learning is superior over the one based on single kernel in the performance of PR curve. When classify the images in database into seven categories, the accuracy of image classification improved eighteen percent than the single kernel model. On the basis of the paper's algorithm, a website of image classification based on multiple kernel learning and image visual feature is built. The website is developed based on J2 EE framework, which makes the function of classifying images uploaded by users come true. Compared with the traditional image retrieval station, the designed website allows users to upload more images one-time, which satisfies users' demands for the classification of local images and finds applications in various scenarios such as image management in social network, images classification in web portals.