Research On Clustering Of Natural Scene Images With Texts Based On Random Projection

Texts in natural scene images are key clues to describe and understand the contents of the scene. They are great valuable for the scene of the expression and understanding of visual information. Because of the complicate background of image texts, large geometrical deformation resulted from different perspectives, the inevitable illumination changes and non-uniform character color, using traditional segmentation methods has many limitations.This paper studies technical difficulties of separation of the scene image background and character recognition. Because of the special of text area, color, texture or other global feature has weak describe ability. This paper proposed using SIFT and Affine-SIFT local feature descriptor methods to extract the image features. The number of key points is too much through those two methods which will be a great influence for post-processing efficiency so this paper filters the features by combination of the main direction and the auxiliary direction in SIFT algorithm and in Affine-SIFT method gets the effective features by matching the image itself.Because of different images, the number of detected feature points is also different. Set of feature points can not directly evaluate the degree of similarity between images. This paper proposes a feature mapping method based on random projection. This method uses image local feature description as input to obtain the image feature vector which ensures the feasibility of the similarity measure between images. Comparing with the feature point matching method, its efficiency is much faster.According different data sets, this paper uses K-means and Affine- propagation methods to cluster and compares the experiment results. Experimental results show that the method can be effective clustering for the natural scene images that contain text characters and the accuracy is 86.66%.