FTNet: A fast trilateral network for surface defect detection in mobile phone

The pre-delivery detection of defects on smartphone screens is critical to user experience and device functionality. These defects, particularly minor scratches, are often small relative to the total screen area. Consequently, the development of a detection system capable of accurately localizing these defects while performing rapid inference presents a significant challenge. This paper proposes a novel, fast, and accurate network, the Fast Trilateral Network (FTNet), which comprises three branches: a context branch, a low-level detail branch, and a high-level detail branch. Two detail branches of varying depths are employed to capture fine-grained details and prevent the omission of small defects. A Multi-Scale Fusion Block (MSFB) is proposed to progressively extract high-level semantic features and effectively integrate them with features from the current stage. This fusion of multi-scale information allows high-level features to be leveraged at an early stage to guide the processing of local details. The MSFBs are stacked to form the contextual branch of FTNet, designated for semantic feature extraction. Furthermore, the Parallel Aggregation Pyramid Pooling Module (PAPPM) is utilized to expand the receptive field and aggregate multi-scale contextual information. Finally, a simplified Feature Fusion Module (FFM) is employed to fuse the feature outputs from the three distinct branches. The performance of FTNet was evaluated on three datasets, and its practical application was demonstrated in industrial smartphone screen defect detection. In comparison with other state-of-the-art (SOTA) models, FTNet demonstrates a superior trade-off between speed and accuracy, achieving 89.02% mIoU at 82.11 FPS.