AEGLR-Net: Attention enhanced global–local refined network for accurate detection of car body surface defects

•An AEGLR-Net is proposed to detect defects on car body surfaces, addressing the significant challenge posed by complex backgrounds to automated defect detection.•An adaptive Transformer–CNN tandem backbone is proposed to capture local features and adaptively perceive global information.•A refined cross-dimensional aggregation attention is proposed to enhance the representation of defects.•An attention-embedded flexible feature pyramid network is designed to guide the network for targeted feature fusion.•Comparative experiments demonstrate that the AEGLR-Net outperforms state-of-the-art approaches. The complex background on the car body surface, such as the orange peel-like texture and shiny metallic powder, poses a considerable challenge to automated defect detection. Two mainstream methods are currently used to tackle this challenge: global information-based and attention mechanism-based methods. However, these methods lack the capability to integrate valuable global-to-local information and explore deeper distinguishable features, thereby affecting the overall detection performance. To address this issue, we propose a novel attention enhanced global–local refined detection network (AEGLR-Net), which can perform effective global-to-local refined feature extraction and fusion. First, we design an adaptive Transformer–CNN tandem backbone (ATCT-backbone) to dynamically aware valuable global information and integrate local details to comprehensively extract specific features between defects and complex backgrounds. Then, we propose a novel refined cross-dimensional aggregation (RCDA) attention to facilitate the point-to-point interaction of multidimensional information, effectively emphasizing the representation of deeper discriminative defect features. Finally, we construct an attention-embedded flexible feature pyramid network (AE-FFPN), which incorporates the RCDA attention to guide the feature pyramid network in targeted feature fusion, thereby enhancing the efficiency of feature fusion in the detection model. Extensive comparative experiments demonstrate that the AEGLR-Net outperforms state-of-the-art approaches, attaining exceptional performance with 89.2 % mAP (mean average precision) and 85.5 FPS (frames per second).