Feature-reinforced dual-encoder aggregation network for flue-cured tobacco grading

Abstract

Flue-cured tobacco grading plays a vital role in product acquisition and planting management. Due to the appearance similarity of inter-class tobacco leaves, the grading accuracy of existing algorithms could not conform to the requirements of accurate acquisition pricing and refined planting management. In this paper, we propose a feature-reinforced dual-encoder aggregation network (FDANet) on a machine vision-based platform for reliable tobacco grading. First, we propose a parallel dual-encoder structure consisting of the main encoder MobileNetV2 and the auxiliary encoder Swin Transformer, which is an indispensable component in the proposed FDANet. Furthermore, a well-designed multi-pooling channel aggregation (MCA) module extracts channel-related global information from the auxiliary encoder, which is leveraged to guide the main encoder. MCA transmits local effective information of the parallel dual-encoder to improve the classification accuracy. Additionally, the SMobileNetV2 blocks that embed the soft-attention module into the original MobileNetV2 blocks are proposed to enhance the representation of key channels. Finally, a feature reinforcement architecture is proposed to integrate the thickness of main vein into visual feature vector since it is a crucial feature perceived by the human visual system. The experimental results show that the proposed FDANet achieves mean test Accuracy of 79.30%, outperforming the novel Swin Transformer by 6.34%. In addition, our FDANet brings significant improvement compared to six state-of-the-art classification algorithms.