Abstract
To address the low recognition accuracy of high-frequency workpiece images caused by complex intra-class textures and minor inter-class differences in top-surface features, we propose a Multi-Branch EfficientNet (MBEN) algorithm for recognizing fine-grained high-frequency workpieces. First, a weakly supervised region detection module is used to obtain discriminative regional images of the workpieces, which are then combined with global images to construct a multi-branch network that enhances the model’s multi-scale representational capacity. Next, by incorporating a weight-adjustment mechanism, we implement joint supervision using an adaptive cross-entropy loss and an adversarial center loss to guide the model toward intra-class compactness and inter-class separability of workpiece features. Finally, a branch fusion module is employed to augment the network’s attention to both global and local information, yielding improved fine-grained recognition performance. Experimental results demonstrate that the proposed algorithm effectively discriminates fine-grained high-frequency workpieces and outperforms existing models and methods in recognition accuracy, achieving an accuracy of 98.75%.