Dynamic real-time detection for corn kernel breakage rate based on deep learning and sliding window technology

Abstract

Currently, the field of intelligent corn harvesting in China lacks effective methods for detecting corn kernel breakage. This paper explores and proposes a corn kernel detection technology that utilizes deep learning and sliding window technology, combined with a specially developed quantitative model, to enable real-time detection of the corn kernel breakage rate. In this study, we quantified the corn kernel mass at various levels of crushing and proposed a quantitative model for the corn kernel breakage rate, which is suitable for real-time computation by a computer vision system. We developed a specialized corn kernel detection device to generate high-quality datasets and retrain our previously proposed corn kernel breakage detection model (BCK-YOLOv7). Subsequently, ablation experiments were conducted to assess the generalization capability of the BCK-YOLOv7 model in corn kernel detection. Furthermore, we analyzed the limitations of single-frame detection through dynamic comparison experiments. To address the instability of single-frame detection results in the corn kernels flow state, we introduced the sliding window technique, which, along with pipeline technology, significantly enhances detection efficiency. Finally, the comprehensive performance of the proposed corn kernel breakage detection technology was validated through systematic testing. The results indicate that the relative error in the detection of the breakage rate remains around 7%, and the detection rate of the technology, when deployed on edge devices, can achieve 22 frames per second (FPS), thereby meeting the requirements for real-time detection of corn kernel breakage rate.