Based On The Feature Of Middle Level In Fine-grained Image Categorization

In recent days,object classification and detection in image processing are getting more and more attention and research.Object classification include two sides,one is basic classification,another is fine-grained classification which attracts more and more people to study.Basic classification is to classify the object which have the different shapes and visual appearance(e.g.: for a bird picture,the basic classification is to identify the object in the picture is a bird or dog).Fine-grained classification is to classification the object with the similar shapes and visual appearance(e.g.: for a bird picture,the fine-grained classification is to identify which kind of birds it belongs to).In this paper,we study the problem of fine-grained image categorization,which is much more useful in real applications than the basic image classification.The whole study in this thesis is based on the most challenge dataset,CUB-200,Stanford_Dogs_Dataset.Instead of low level descriptor as sift and hog feature,we adopt the BoW frame work and efficient match kernel(EMK)technique to map the low level descriptors into the middle level representations.Combined with the weighted spatial pyramid,which can compensate the loss of spacial information in both BoW and EMK framework,it can achieve state-of-art performance.Compared with BoW,which can also be viewed as kernel matching approach,EMK digs the relations among vocabulary bases and finds a new mapping in kernel framework.By it,local features are mapped to a low dimensional feature space and average the resulting vectors to form a set level feature in EMK.It is proved that it is helpful to improve the system performance.