Data Augmentation Based on Free-Form Deformation for Conductive Particles Detection

False positives and missed detections for over-pressured and under-pressured particles in TFT-LCD quality inspection could happen in existing machine learning-based methods. The main reason for the poor detection is the seriously unbalanced data between strong and weak particles. Therefore, a data augmentation method based on free-form deformation for conductive particles detection is proposed. The image is converted into a three-dimensional point cloud in this method. The method takes advantage of the characteristic of the simple operation and
$G^{k}$
continuity of the free-form deformation to change the coordinates of the three-dimensional point cloud of conductive particles. The original geometric characteristics of the conductive particles are maintained at the same time. Thereby the amount of over-pressured and under-pressured particles is increased. Compared with commonly used data augmentation methods based on geometric transformations such as mirroring and shearing, the method can better maintain the original geometric characteristics of conductive particles and achieve better detection results. In the experiment, only more than 3000 conductive particle images obtained by intensity deformation from more than 300 artificially annotated conductive particle images were used. The detection rate of over-pressured and under-pressured particles is increased from 80% to 97% using the trained model based on Mask R-CNN.