Conductive particle detection via efficient encoder–decoder network

Particle detection aims to accurately locate and count valid particles in pad images to ensure the performance of electrical connections in the chip-on-glass (COG) process. However, existing methods fail to achieve both high detection accuracy and inference efficiency in real applications. To solve this problem, we design a computation-efficient particle detection network (PAD-Net) with an encoder–decoder architecture, making a good trade-off between inference efficiency and accuracy. In the encoder part, MobileNetV3 is tailored to greatly reduce parameters at a little cost of accuracy drop. And the decoder part is designed by using the light-weight RefineNet, which can further boost particle detection performance. Besides, the proposed network adopts the adaptive attention loss (termed AAL), which improves the detection accuracy with a negligible increase in computation cost. Finally, we employ a knowledge distillation strategy to further enhance the final detection performance without increasing the parameters and floating-point operations (FLOPs) of PAD-Net. Experimental results on the real datasets demonstrate that our methods can achieve high-accuracy and real-time detection performance on valid particles compared with the state-of-the-art methods.