A Feature Prefusion and Mask-Guided Network for Camera Decoration Defect Detection

Camera decoration is an important part of smartphone. To achieve fully automated production, a dependable, efficient, and automatic method is required for camera decoration surface defect detection. This article presents a detection scheme based on computer vision to improve the efficiency of screening defective products. Since there is no available dataset for method designing in camera decoration field, we establish a camera decoration defect dataset CD3 including 9417 samples with four types of defects. To increase sample size and alleviate category imbalance of CD3, we provide a dataset enhancing framework including a defect copy method and a background reuse method to generate an enhanced dataset CD3_En containing 39649 samples. Besides, a feature fusion and mask-guided network (FMN) including a feature prefusion (FPF) module and a multistage fusion (MSF) module is proposed to screen the defective products. The FPF is constructed by receptive field blocks (RFBs) and information diffusions (IDs), and it can achieve data volume reduction and context enhancement after being embedded between the BoneNet and Neck. The MSF is used as the Neck to realize a two-step feature fusion for predicting the bounding boxes of defects and their masks. The experimental results on the CD3_En dataset demonstrate the superiority of the proposed method compared with other 11 classic object detection methods.