Natural Image Classification Method Based On The Multitasking Learning Research

Along with the computer and the Internet widely popularity, people more and more contact and use natural images, natural images are growing so fast that people almost no time to analysis and tidy up these natural images. However the current work of image classification mainly focused on medicine, remote sensing and other professional image fields, yet there have a small study of natural images classification method, which is mainly traditional single task learning. This mechanism has poor generalization ability. This paper aims to use the related approaches in data mining and pattern recognition domain to study the problem of natural image’s classification, tries to use multitasking learning mechanism to improve generalization capability of classifier. In order to achieve this goal, this paper made the following jobs:1. This paper analyzed the characteristics of the natural images and the present situation of natural images’ classification. A method that fills the gap between low visual feature space and high-level sematic was proposed. And then achieved the purpose of efficient retrieval;2. Kinds of methods that extract the natural image’s feature in low visual space were studied. Image’s color features was extracted including color histogram, color space torque and color polymerization vector. Texture features was extracted including Tamura texture, gray symbiotic matrix and Gabor filtering texture. In order to get more reasonable features, we compared all these features’ expression ability. Finally, we combined color histogram that was weighted by different coefficient and Gabor texture to form a112dimensions comprehensive feature. These high dimension features were shown in parallel coordinates, which give us some useful information about whether all the images can be divided.3. This paper studied several classification methods for data in pattern recognition, including SVM (support vector machine), neural network, bayesian classifier and k-the most neighbor classifier, analyzed their fitness in natural images’classification;4. A method that measures the relatedness between any tow sample vectors using the linear correlation of tow any sample vectors was proposed. And then make a cluster for all the tasks according to the relatedness. Which tasks should be put together to learn was decided by the relatedness. We finished the classification of natural images using multitask learning, and proved the natural images’classification’s generalization. Finally, we showed multi-task learning’s effectiveness in natural images’classification through some experiments.