Date Imprint Extraction And Recognition From Color Photographs

Date imprint in natural scene photographs is one of valuable information for indexing and retrieval. Date imprint in photographs is a kind of text in video or in image. In recent years, how to index and retrieve text in video or image is a research hotspot in the field of Content Based Video Retrieval (CBVR). There are two kinds of text in video:scene text and artificial text and date imprint in photos is a kind of artificial text.Differing from the traditional text in video or image, in general, date imprint has not sharp edges, uniform stroke color or definite intervals between adjacent characters. In this case, the common methods such as color-clustering based algorithms and edge or feature based algorithms are not as efficient for video or image text as for date imprint.In this thesis, we propose a novel two-step color-based approach to detect the .date. The R component of the photo image is enhanced to the R1 component so that four potential regions can be found in the four original corner areas. Then the smoothed histogram by Gaussian filter of another enhanced R component R2 is analyzed to find the last peak. Those pixels contained in this peak can be seen as the date pixels. A proper valley in the left side of the last peak can be found to represent the binary threshold. After extracting the edge images of the four potential regions, the binary images are combined with the edge images to new images by color-edge-based method. On those new images, the coarse-to-fine method and heuristic information are used to locate the date string accurately. On the thinned image of the located region, the date string can be segmented into characters which are ready to be input to the module of recognition. Finally, the number of the strokes of a character at different positions are calculated to build a feature vector of 15 dimensions. A new date character can be recognized by the rule of Minimum Weighted Euclidean Distance.Experiments show that our approaches can effectively extract and recognize the date characters. This thesis provides a rudimental research in this new field of photo date imprint detection and recognition.