Detecting Conductive Particles in TFT-LCD with U-MultiNet

The conductivity of the TFT-LCD circuit can be determined by counting and locating anisotropic conductive film (ACF) particles in the circuit, which is a critical step for the detection of conductivity in the TFT-LCD manufacturing process. In order to solve the aggregation and overlap problems of ACF particles, in this paper, we propose a U-shaped multi-scale convolution network(U-MultiNet). The U-Net network structure is designed adaptively, which reduces the parameter amount of the network. The multi-scale convolution blocks are introduced for extracting the multiscale spatial features. In addition, the particles detection is transformed into a segmentation task, which helps to solve the aggregation and overlap problems of particles. The experimental results show that the method achieves high precision and recall rate, which is far superior to the previous methods.