Image Automatic Annotation Methods Based On Conditional Random Field And Error Correcting Output Code

The image data becomes much more easily accessible to people and usually carries more information than text, thus the image becomes increasingly widespread with the development of the World Wide Web and multimedia technology. A critical issue now is how to retrieve information which users are interested effectively among vast amount of data. At this time, Content-based image retrieval technology arises and becomes a hot problem in the image research field. The Image Automatic Annotation technology is the key to image retrieval and image intelligent recognition, the fundamental job of object recognition and image understanding. Image Automatic Annotation technology uses some keywords or image content to label image automatically, those labels can describe the Semantic content of image and are very suitable under the background of Large-scale image retrieval. This paper introduced the Conditional Random Field(CRF) model to the research of image labeling. The CRF model incorporates neighborhood interactions in the labels and observed data, thus it has many advantages over traditional generative models. Also the CRF model has the disadvantages such as the model complexity and time consuming progress of model training. Based on the above reasons, this paper combined the Error Correcting Output Code with image labeling, this model is simple and easy to build, which can reduce the time complexity. The major work from this paper is listed below:1. Digital image stored by means of independent pixels, the large amount of data and redundant information have been a intractable problem in the image processing field. To reduce the time complexity of subsequent image algorithms, this paper over-segment the image into superpixels based on correlation between pixels. The follow-up algorithms are built on superpixels instead of pixels.2. In probabilistic graph model, superpixels represent the node of a graph, we model the relationship between variables based on the CRF, the test image can be labeled automatically after the parameters of model are learned during the process of training. The CRF model take the advantage of the ability that can better illustration the interaction between superpixels labels.3. Given the fact that CRF needs labels for all hidden states for the whole training set, and it is difficult and time consuming. This paper novelly combined the ECOC technology with Image Automatic Annotation, converts the multiple classification problem into binary classification problem based on the thought of coding, which greatly reduce the time complexity.