Deep learning based soybean seed classification

Abstract

Accurately sorting high-quality soybean seeds is a crucial and time-consuming task in quality inspection and food safety. This paper designs a full pipeline to classify the soybean seeds, which follows a segmentation–classification procedure. The image segmentation is performed by a popular deep learning method, the Mask R-CNN, while the classification stage is performed through a novel network, named Soybean Network (SNet). SNet is an extremely lightweight model based on convolutional networks, and it contains mixed feature recalibration (MFR) modules. The MFR module is designed to improve the representation ability of our SNet for damage features so that the model pays more attention to the key regions. Experimental results show that the proposed SNet model achieves 96.2% identification accuracy with only 1.29M parameters, which outperforms six previous state-of-the-art models. The proposed SNet could be used for the automatic recognition of soybean seeds on the resource-limited platform.