The Study On Image Classification

With the rapid development of computer and Internet technology, we have entered the image era. However, the explosive growth of web image data induces an awkward situation-rich information and poor knowledge. On the one hand, large amount of image information are available on the web site; on the other hand, the data that people need is hard to obtain. In order to meet needs of retrieval, people deeply hope the computer could automatically analyze and understand images. Image classification technology is the effective way to solve this problem, because it can build the semantic information of image automatically consistent with human cognition.Due to the ambiguity of image, the multiple instance learning framework, with its excellent image representation capability, was successfully applied to image classification task. Image classification method based on multi-instance learning can be roughly divided into two directions:one is making the algorithm adapt to the data, that is, under the multi-instance framework to study new algorithms; another idea is the data fit algorithm, converting the multi-instance learning problem into the mature traditional supervised learning problem. Both methods have achieved good performance.Based on the degradation strategy, we also proposed a novel multi instance learning algorithm to address image classification task. First, a new instance prototype extraction algorithm was proposed to obtain instance prototypes for keywords. And then, x-means algorithm was adopted to cluster instance prototypes for constructing projection space, this step transformed the MIL task into a traditional supervised learning task. Finally, an SVM was trained for classification decision.We have tested our proposed algorithm on the benchmark Corel5k data set. Experimental results showed that our new instance prototype extraction algorithm can result in more reliable instance prototypes rapidly with less noisy; moreover, comparison with some state-of-the-art multi instance learning approaches, the proposed approach obtained better results in the image classification task with higher accuracy, which verified the effectiveness of our algorithm.