Real-Time Detection of Impurity and Crushing Rates in Machine Harvested Soybean Kernels Using Scse-Unet

Abstract

Soybean impurity rate and crushing rate are critical performance indicators for evaluating the efficiency of soybean combined harvesting machinery. Rapid and accurate acquisition of these metrics during the harvesting process is essential for achieving intelligent soybean harvesting. This study proposes a deep learning-based detection method to address this challenge. An image acquisition device was designed to capture high-quality images of soybeans laid flat on a conveyor belt, creating a robust dataset for analysis. To tackle the issues of semantic complexity and image adhesion in the acquired images, we introduce the scSE-UNet network model. This model enhances the traditional U-Net architecture by replacing its encoder with the VGG16 structure and integrating scSE-block+ at the end of the decoder's skip connection layer. These modifications enable the model to automatically learn effective features from both spatial and channel dimensions, significantly improving its ability to extract key features. Additionally, a hybrid loss function combining Dice Loss and CE Loss is adopted to mitigate accuracy degradation caused by the severe imbalance in pixel distribution. Experimental results demonstrate that the improved scSE-UNet model achieves effective segmentation and classification of whole soybeans, broken soybeans, and impurities, with F1 scores of 97.8%, 97.2%, and 96.3%, respectively, and a mean Intersection over Union (mIoU) of 93.9%. The average detection time for a single image is 0.153 seconds, meeting the requirements for real-time accurate detection. This study provides a reliable reference for the efficient and precise detection of impurity and crushing rates in soybean harvesting.

Keywords: image segmentation, Soybean, Impurity Rate, Crushing Rate, Deep learning, real-time detection