A Weight‐Complementary Confusing Background Network for Industrial Defect Detection

ABSTRACT Defect detection is a critical component of industrial production, as the timely and accurate identification of defects is paramount to ensuring product quality and enhancing manufacturing efficiency. However, extant methodologies continue to confront specific challenges in the context of targets that are analogous to the background. Most methods tend to lose accuracy when focusing on spatial features during detection. The edge interference issue engenders a more arduous environment for feature extraction. This paper puts forth a novel edge‐aware target detection architecture, designated as FL‐DFINE, which is based on the DFINE network. The objective of this architecture is to enhance the detection capability of the model in complex industrial environments. An efficient cross‐layer multi‐scale fusion module (CMF) is designed to realize effective feature alignment between different semantic layers by introducing an interlayer weight matching mechanism. Second, an enhanced feature transfer backbone network, L‐HGNetV2, is proposed to ensure that high‐quality features can be stably transmitted to the deep network. The consistent descriptor weighting (CDW) module is designed to capture multi‐scale contextual information through convolutional operations at varying expansion rates. A systematic evaluation of FL‐DFINE was conducted on three industrial defect detection datasets, which were constructed for industrial vision inspection tasks and are highly representative. A substantial body of experimental evidence demonstrates the superior performance of FL‐DFINE in comparison to prevailing mainstream methods across multiple metrics, including AP50, APS, APM, and APL. Its efficacy is evident not only in its enhanced detection accuracy but also in its capacity to maintain equilibrium across diverse target scales. Code address: https://github.com/chaojiaidanya/FLDFINE .