Conditional Random Field Based Object Extraction

Object extraction is the process of separating interested targets from background. It is an important part of the computer vision and a key step of image understanding and recognition. Accurate object extraction is still a challenging task due to the complexity of object and variability of background. Relations between the current location and its surrounding area can effectively reduce the negative impacts of the uncertainty and ambiguity of the image on object extraction. Hence, how to use the context information becomes a popular issue of object extraction.Conditional Random Fields (CRF), developed from Markov Random Field (MRF), can instruct the local judgments by not only the connections of the adjacent regions but also the information of the whole observation field to extract object regions more reasonably. This paper focuses on the CRF-based object extraction methods from two aspects. On the one hand, the suitable features for extraction are extracted by analyzing the characteristics of objects. On the other hand, the CRF-based object extraction framework is modified to take full advantage of the relationship among the objects. The main contributions of this paper are as follows:A fast CRF model inference method is proposed. Model inference is the process of obtaining the optimal target label by using the trained CRF model. The inference time increases dramatically as the increase of image size. First, the proposed method infers on a low resolution image, which effectively reduces the convergence time due to the small number of pixels. However, the extracted target is rather coarse. Then, the model inference is applied again on the edge region of the object, based on the inference result on low resolution image, to obtain a relatively fine target. The proposed method effectively shortens the CRF model inference time in case of not reducing the accuracy of target extraction.An object extraction method which uses CRF to combine different scales and directions of contour fragments is proposed. The contour is one of the most obvious features to distinguish between the targets and background. Decomposing the contour into fragments will be more insensitive to the deformation. The contour feature is expanded to a variety of scales to detect objects with different sizes. Then, the location of the candidate fragments are detected by the hinge angle, the contour direction, as well as partial Hausdorff distance. CRF combines the candidate fragments with different scales and directions and efficiently uses their relations to select the final contour.A natural scene text extraction method based on CRF with global feature is proposed. The candidate character regions are extracted by Toggle Mapping combined with an edge filter, which can extract the candidate regions with low contrast or noise more efficiently. Local features such as size, color and texture are not suitable to represent the text due to their variances, while the similarities between the current node and its neighboring nodes are more stable to be global features. These global features are combined by CRF to effectively extract the text regions.A text extraction method for document image based on a two-layer CRF is proposed. With the real or imaginary part of Gabor, text regions have strong filtering results which make the text regions be different from background. The image is divided into grids with same size and CRF is used to distinguish the text from background with the histogram taken from the filtering results of the neighborhood of each grid. In order to optimize the extraction results, a two-layer CRF model is applied to compromise the classification results of the two characteristics and further improve the accuracy of the text extraction.A handwritten character extraction method, which combines CRF with Support Vector Machine, is proposed. First, a dual-threshold binarization method based on Toggle Mapping is introduced to extract the characters in the document image with uneven illumination. The input image is divided into grids with same size to avoid the case of direct handling the adhesion of handwritten and printed characters. Then, the characteristic Edge Co-occurrence Matrix is extracted from the neighborhood of each grid. Due to the similarities of adjacent grids, CRF classification framework is employed to divide the grids into the categories of the handwritten and the printed. CRF classification framework combined with Support Vector Machine can obtain more reasonable classification results. Finally, these results, using the post-process of text line information, will be more precise and meaningful.