Two-Stage Visual Detection of Micro-Objects: Theory and Experiments

As one of few effective perception measures in the micro world, vision provides important feedback about the position of micro-objects; Different from the visual perception at the macro scale, the perception in the micro world is facing its own challenges. For example, multiple micro-objects often sway in the medium or overlap at different depths; The environment is dynamic in the sense that some objects may enter or leave the field of view during the manipulation. This paper reports the development of a deep vision network for micromanipulation, which is able to detect micro-objects and output its real-time position with high accuracy and robustness. In addition, a position correction algorithm is proposed to calculate the position offset through the pre-sequence time series, to further improve the position information generated by the network. Experimental results are presented to illustrate the effectiveness of the proposed method, that is, small tracking errors have been obtained for the detection of micro-objects; Moreover, such detection results are also with high success rate and stability for the micromanipulator.