Study On Local Invariant Features In Image Classification

Image classification is an importan t access to the category of objects(cats, cars, etc.) and scene types(beach, vegetation, etc.). It has been widely used in mult imedia retrieval, vision-based navigation, intelligent surveillance and other field s. However, images always have v arious geometric and photometric changes, so that m ake it difficult to accurately extract the visual information in images. If the process of extracting image visual information can not cope with these changes, it is unable to exactly classify the images in different categories. The local invariant features have not only excellent distinctiveness, but also great invariance in a variety of geometric and photometric changes. Therefore, it is crucial in an image classification system to study local invariant features.Given great success Bag-of-Words(Bo W) model have made in applications of image classification, in the process of image classification experiments in this paper images were represented by Bow for global description. The basic framework and core technology of Bo W was presented in the second chapter. The fast k-means algorithm was used to construct visual dictionary to accelerate the process of experiment. And the pyramid space was used to join space location information of local invariant features to word Bo W model. In this thesis, we studied the sampling and description of local invariant features aim to dispose changes(scale, rotation changes, photometric changes) which is the most most common and influential problem in image classification. We proposed the Rotation Tranformed Features(RTF) on the basis of fast dense SIFT. The RTF were computed with specific sampling rules and described in an approximate way of color SIFT, thus distinctive features with photometric invariance were computed with a fast method. And, features were tranformed in an equivalent way of multi-orientation sampling, thus it can improve the rotation invariance with little cost. Then, the pyramid sampling is used to increase the scale invariance in sampling. The effectiveness and efficie ncy of RTF in image classification was verified through experiments.In the fifth charpter, local invariant features was applied in the rotary kiln sintering state recognition in the rotary kiln. Flame is a kind of radiant object and the brightness and color of the flame suggest the state of clinker, so the photometric invariant features could not conducive to accurate description of the image of rotary kiln. We designed the Local Shape and Color Features use the characteristics of the flame image of rotary kiln, then it was used for the recognition of the rotary kiln sintering state and obtained a high precision.