The Defect Detection Algorithm for Sintered NdFeB Based on ADSEAM-YOLO

To address the challenges commonly encountered in the detection of sintered NdFeB, such as insufficient speed and accuracy, difficulty in detecting small targets, and noise interference, this paper proposes an innovative and efficient algorithm framework. This framework is based on an Attention-guided Convolutional Neural Network and a Separation and Enhancement Attention Module, referred to as ADSEAM-YOLO. Firstly, to effectively tackle noise interference, an Attention-guided CNN is integrated into the Backbone network. By combining the strengths of both dilated convolution and conventional convolution, this design enhances the network's ability to capture key features while significantly reducing the impact of background noise, thus laying a solid foundation for subsequent accurate detection. Secondly, at the critical feature fusion stage (Neck layer), the SEAM module is introduced. This module fully leverages multi-scale feature information and employs depthwise separable convolution to deeply explore and learn the complex correlations between spatial dimensions and channels. This innovation not only greatly improves the algorithm's ability to detect micro-defects in sintered NdFeB but also ensures the comprehensiveness and accuracy of the detection results. Additionally, to further optimize model performance, the DualConv module is designed by cleverly combining 3×3 group convolution and 1×1 pointwise convolution. This design solves the issue of poor inter-channel communication while effectively preserving critical information from the original input feature map. As a result, it enhances detection accuracy while achieving model lightweighting, thus improving processing speed and efficiency. The experimental results show that, on the sintered NdFeB dataset, the model achieves an improvement of 7.5% in [email protected] and 2.5% in [email protected]:0.95 compared to the baseline model. With a model size of 52.3 MB and 126.3 GFLOPs, it maintains the lightweight parameters of the baseline model while significantly enhancing detection performance.