Design of a Detection and Sorting System for Broken Corn Kernels with an Online Identification Method

Highlights A novel online detection and sorting system for broken corn kernels has been developed. A corn kernels image processing method based on HSV space color threshold segmentation was proposed. A magnetic suction openable workstation, a limit switch, and an opening and closing cover paddle were designed for the sorting mechanism. GoogleNet was determined to be the optimal model after comparing the three classic network models of AlexNet, GoogleNet, and ResNet34. Abstract. During threshing and harvesting processes, corn is exposed to compression force, impact, and collisions, which collectively contribute to kernel breakage. To overcome the limitations of low accuracy and high cost inherent in manual sorting, a novel detection and sorting system for broken corn kernels has been developed. This system combines a seed discharge unit, a sorting unit, a control system, and image recognition technology to enhance sorting precision. A Raspberry Pi serves as the host computer, while an STM32 microcontroller serves as the subordinate unit, responsible for detecting the status of limit switches. Furthermore, mechanical analysis and innovative design enhancements were performed on key components of the detection system. To improve recognition accuracy, a corn kernel image processing method based on spatial color threshold segmentation was developed. AlexNet, GoogleNet, and ResNet34 were utilized to train six datasets of varying magnitudes. By loading 18 weight files, the optimal image scale that achieved the highest accuracy was identified. The recognition accuracy of the three network models was compared across the validation set, test set, and broken corn kernels. Results indicated that the GoogleNet network model demonstrated the best overall performance, achieving an accuracy of 95.58% on the validation set and 94.55% on the test set. In evaluating the detection device, the GoogleNet model achieved accuracies of 92.00%, 95.00%, and 98.00% for identifying broken, moldy, and intact corn kernels, respectively, with a sorting accuracy of 90.00% specifically for broken kernels. Keywords: Broken, Convolutional neural network, Corn kernels, Detection, Sorting.