Research On Key Techniques Of Image Representation In Recognition

Object and scene recognition has wide applications in content-based image retrieval, Autonomous robot , medical image classification etc. We reviewed the major progress these years, and then explored some key technology in computer vision based recognition, in a bid to further increase the recognition accuracy.Classical object/scene recognition framework can be divided into two parts: image representation and image classification; the former tries to get the most discriminative information on the meaning of the object expression; while the latter one try to separate different categories by training on specific image set.Similarly, this framework also applies to self-organization recognition in videos. However, there still exist unique characteristics:? Characteristics of video goal often experience long-term gradual process and therefore its features are bound to experience this change process. This requires the analysis of the effectiveness of the feature must be a progressive process.? Object appeared accompanies with scenes, that means the target and the background has a strong correlation. How to use this correlation to improve recognition performance, is also big challenge.However, there exist huge gap of the performance between existing methods with human object recognition capacity.There is no evidence that human pattern recognition algorithm is superior than the standard machine learning algorithms, and human doesn't dependent on the amount of training data size very much. Therefore the reason affecting the accuracy of recognition may be the choice of features.In fact, feature description plays a more important role than others. Thus, we focus research on how to describe the characteristics of video goal effectively. On one hand, the target properties of gradient requires establishing features online evaluation system: specific features may be effective only in a specific time period; on the other hand, the relevance of objectives and scenarios can be achieved by mixing global scene expression with local features.We started from the bag of features model, take the high-dimensional features as the superposition of one-dimensional feature,then one-dimensional features has unknown probability distribution of the observation. Therefore, the online evaluation of features turned into the on line estimation and comparison of probability density. The idea of Monte Carlo is referenced and gaussian mixture model is used to appropriate the probability distribution, KL divergence is used as the main measure, feature robustness and the role for decision are analysised from the perspective of mutual information, leading to online feature evaluation model preliminary.