De Chen, Wenbo Qiu, Zhiwen Zeng, Xuexian Li, Qingdong Yan, Qin Zou
Surface defect detection, GANs, RL, multi-scale feature fusion,atrous convolution
Surface defect detection is crucial for quality control in lithium battery cell manufacturing. Due to the scarcity of defective samples, existing methods typically rely on generative adversarial networks (GAN) to generate defective samples and corresponding pseudo- labels, followed by the design of a deep defect detection network. This allows for the training of a competent defect detection model even with a shortage of defective samples. However, GAN models still face stability issues such as mode collapse. To address this, this study proposes a novel end-to-end trainable network that innovatively applies reinforcement learning algorithms to optimise the generation strategy of GAN. A reward mechanism is designed between the discriminator’s output and the generator’s production, ensuring robust changes in the GAN training gradient. Experiments demonstrate that our proposed method has made significant progress in generating samples that are closer to the true data distribution. In addition, this study designs a multi-scale feature fusion and atrous convolution module to integrate multi-scale information, enhancing the detection capability of subtle defects. In test scenarios aimed at detecting small target defects, the model proposed by this research achieves state-of-the-art precision in segmenting small target defects on the surface of battery cells.
Important Links:
Go Back
