Deep learning-based machine vision algorithm for target detection and positional solution of ultrathin vapour chamber

author: Haoting Han, Yong Li, Wenjie Zhou, and Zhaoshu Chen
status: in the second round of peer review
abstract:
The fast and precise positioning of ultrathin vapour chamber (UTVC) is essential to realize the automated liquid injection and vacuuming processes. In order to acquire the position information of UTVC in the material box rapidly and accurately, a positioning algorithm based on Mask R-CNN and improved minimum bounding rectangle (IMBR) is proposed to correctly detect and localize the UTVC of arbitrary shape (consisting of circular arcs and straight lines) for automatic and accurate liquid injection and vacuuming processes. Firstly, Mask R-CNN is applied to extract the mask of the liquid injection port feature of UTVC, and the centroid and centerlinear equation of the mask are obtained by using the sub-pixel contour extraction and MBR combined with the dimensional feature of the liquid injection port. Then, image preprocessing is performed on the UTVC image to screen out the cover region, and linear equation over the centroid of the cover region is obtained by using the sub-pixel contour extraction and MBR combined with the centerlinear equation of the mask. Finally, the intersection point of the two straight lines is derived as the positioning point of the UTVC, and the direction vector from the positioning point to the centroid of the mask is constructed to solve the orientation angle of the UTVC, so as to make up for the lack of directional judgment of the MBR through the direction vector, and thus to complete the improvement of solving the orientation angle of the MBR. Experiments show that the maximum positioning point error and the maximum orientation angle error of the algorithm are 0.092178mm and 0.26827°, respectively, which meet the actual production requirements.