Real-time segmentation and sorting method for molybdenum ore X-ray images using multi-scale attention SegNet

To address the technical bottlenecks in traditional mineral sorting processes, this paper proposes a SegNet-based industrial visual inspection method for molybdenum ore integrated with a multi-scale attention mechanism. Targeted optimizations are achieved through the construction of a multi-level feature enhancement mechanism: (1) Aiming at the classification confusion caused by the highly similar texture features between low-grade molybdenum ore and tailings, this study proposes integrating a Multi-Scale Adaptive Spatial Attention Gate (MASAG) to enhance the encoder’s ability to capture the texture features specific to molybdenum ore; (2) To address the feature transmission inefficiency in low-contrast and high-density overlapping regions of X-ray images, this study integrates a Convolutional Additive Self-Attention (CAS) module across the encoder–decoder interface, aiming to refine the feature propagation pathway and enhance information interaction between these two components; (3) For the missed detection problem caused by sparse features of small particles after molybdenum sulfide crushing, a Multi-Order Gated Aggregation Network (MogaNet) is integrated to adjust the decoder output, enhance local feature processing, and optimize the segmentation results; (4) To improve the segmentation accuracy of molybdenum ore targets, the hybrid loss of Binary Cross-Entropy and Dice (BCEDiceLoss) is adopted to refine the optimization direction during model training and enhance segmentation accuracy. Validation using a professional dataset of high-resolution molybdenum ore images constructed by X-ray equipment shows that the improved Segmentation Network(SegNet) model achieves comprehensive performance upgrades in detection and segmentation, demonstrating excellent robustness and generalization ability in complex ore scenarios. This delivers reliable technical backing to facilitate the intelligent transformation of mineral processing operations. • A molybdenum ore segmentation method via multi-scale attention SegNet is proposed. • The MASAG module boosts texture capture for low-grade molybdenum ore. • The CAS module optimizes feature transmission between encoder and decoder. • The MogaNet module in the decoder boosts small-particle identification. • The BCEDiceLoss balances pixel/region optimization and refines boundary fitting.