Multi-connected Tiny Object Detection for Color-Strip Recognition
Zhipeng Zhang, Xinya Peng, Xiaohang Yuan, Wenting Ma, Qingchen Liu, Yingteng He, Xingang Chai, Jie Wei, Jinman Lin · 2023
Object detection is one of the most popular ap- plications in AI image recognition, and Multi-connected Tiny Object detection (MTOD) is an important but difficult branch. For example, we call the color bar carrying optical branching device port usage information and device ID information a color strip, where the color bar consists of multiple tiny color blocks with information bits connected to form a large information string, and color-strip recognition is a kind of MTOD. For MTOD, there are two solutions: bottom-up and top-down. We have previously proposed the bottom-up scheme. This paper proposes a top-down scheme—that is, first find a large object linked in a string, and then subdivide it into each unit. Next, the Optical Character Recognition (OCR) method was used to realize the multi-interference multi-connection dense tiny target detec- tion. Furthermore, to reduce the mutual interference between multi-connected objects, we introduce a "guard code" between multi-connected objects to improve the separability between the tiny object. Through experimental verification in real scenes, the proposed scheme achieves very competitive performance in recognizing multiple connected dense tiny objects.