SSGDD-YOLO: Multiscale Feature Fusion and Multiattention-Based YOLO for Smartphone Screen Glass Defect Detection

Abstract

Surface defect detection is essential for ensuring the product quality of smartphone screen glass. In this work, a smartphone screen glass defect detection model based on an enhanced YOLOv7 framework with multiscale feature fusion and multiattention, named SSGDD-you only look once (YOLO) is proposed. In the developed SSGDD-YOLO model, the branch fusion block (BFB) is integrated low-level features from multiple scales through parallel processing, to enhance the details in lower level features for minimizing the information loss as less as possible. Furthermore, the SPPCSPC module of the head is improved as the SPPCSPC-I module, by replacing the standard max pooling with local importance-based pooling (LIP) that reflects the importance of features. The developed SPPCSPC-I module allows the network to automatically learn adaptive importance weights of features during downsampling, enhancing the multiscale feature extraction capability with diverse receptive fields. Finally, a contour-mixed attention block (C-MAB) is inserted into the feature fusion section of the network, which enhances spatial and channel information of features to reduce target information loss, improving the representation capability. Experiments are conducted using a challenging real-world defect image dataset gathered from a smartphone screen glass inspection line in an industrial plant. Results show the proposed SSGDD-YOLO model can achieve the highest mAP of 62.46% among all compared methods.