Image segmentation and object recognition based on bidirectional scanning fusion technique
Hao-peng Xu, Wu-qin Toing, Wei-liang Fan, Jinyu Xiong · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
Nowadays, image segmentation is of great importance in features extraction and object recognition for video image sequence. Many traditional segmentation techniques have special application and exists some limitation in some degree. After analyzing the advantages and disadvantages of the present image segmentation and object recognition methods, according to the characteric of infrared image, this paper proposes a very simple yet effective algorithm to optimize the threshold value, which is in accordance with the status of the reference pixels. The proposed algorithm varies the threshold value with bidirectional line-by-line scanning (the forward scanning and the backward scanning) fusion model. The purpose of the technique is to discriminate targets from the background, which is equivalent to assigning the label "F" (representing "foreground") to object pixels, and the label "B" (representing "background") to background pixels. Based on these bidirectional scanning intersections in the corresponding regions, this paper applies the conditional probability density function (PDF) to fuse and optimize the threshold value. At the same time, the optimal threshold values for target segmentation and recognition were acquired. Therefore, this paper designs a novel background frame differencing method that refers to previously conclusions made by neighboring pixels. Change different infrared image sequences, the experiment results show this fusion method can eliminate the boundaries blurring, especially the transition regions between object (foreground) and background. As a conclusion, for different infrared image sequences with complex illumination change, noise change, etc., the proposed method gives better segmentation and recognition results for objects than other traditional methods, such as the fixed threshold method, the single directional scanning technique, and so on. On the other hand, the proposed method has lower complexity and higher real-time, which is helpful for hardware design and engineering application.