Classification of Unsound Wheat Grains in Terahertz Images Based on Broad Learning System

Wheat quality and safety are crucial for securing the grain supply. The traditional deep learning methods used in the identification and detection of unsound wheat grains often face challenges, such as complex processing steps, computational complexity, and longer data processing times associated with an increase in the depth of the network. To address these issues, this work introduces the concept of a broad learning system (BLS) to develop a terahertz (THz) image detection model for unsound wheat grains detection. This model is based on a radiologist-inspired deep denoising feature pyramid network BLS (R-F-BLS). The radiologist-inspired deep denoising network (RIDNet) denoising module and feature pyramid network (FPN) feature extraction module are embedded in the BLS network structure. The experimental results show that R-F-BLS achieves 94.68% classification accuracy, marking an improvement of 11.98%, 6.8%, 2.66%, and 2.77% over support vector machine (SVM), simple convolutional neural network (simple CNN), Google Inception Net (GoogleNet), and deep residual neural network (ResNet), respectively. Consequently, the R-F-BLS model combines enhanced accuracy with reduced complexity, positioning it as an effective and innovative approach for unsound wheat grain identification.