Bi-attention pyramid network for small defect with complex background in industrial detection

Abstract

Defect detection is essential in modern industrial production for ensuring product quality. However, current methods struggle with small object and multi-scale detection, especially in complex backgrounds, due to limitations in feature extraction and fusion. To address these challenges, this paper introduces the bi-attention pyramid network for enhanced small defect detection in complex backgrounds, inspired by the channel and spatial attention mechanisms, as well as advanced fast single-stage detection neural networks. Firstly, this work develops an innovative bi-attention pyramid network and introduces a new feature fusion module, the bi-attention fusion block (bi-AFB), which to enhance the detection of small defects. Additionally, the integration of dynamic sampling to optimize the upsampling process and the incorporation of a large selective kernel into the backbone network expand the receptive field, thereby improving multi-scale detection and the utilization of background contextual information. Experimental results demonstrate that bi-APNet achieves state-of-the-art performance across three benchmarks, with minimal increases in model parameters, thereby validating the efficiency and superiority of the proposed architecture.