Development of a Deep Learning System for Real-Time Quality Classification Using Multi-View Images

Conventional deep learning-based grading systems for agricultural products typically rely on top-view images captured by conveyor belts, limiting surface inspection. Although random rolling has been introduced to expose different sides, it often results in redundant or inconsistent coverage, reducing classification reliability. This study proposes a multi-view image acquisition and classification system utilizing the structural features of the Calistar grader. Four industrial cameras—two positioned above and two below—captured images from four distinct angles, which were merged into a single composite multiview image. This composite image was processed using a ConvNeXt V2 model, converted to ONNX, and optimized with TensorRT for real-time deployment. The experimental results demonstrated a classification accuracy of approximately 99% and a processing rate of five objects per second. These findings validate the effectiveness of the system in accurately grading agricultural products in real-time and demonstrate its potential applicability to irregularly shaped 3D objects beyond onions.