Real-time canola damage detection: An end-to-end framework with semi-automatic crusher and lightweight ShuffleNetV2_YOLOv5s

Abstract

The current method employed in detecting damages present in canola kernels is inefficient and time-consuming. The sample preparation process is laborious and the judgment between sound and damaged seeds using visual methods is prone to errors as they are small and indistinguishable. To date, there is no hardware and supporting software solution that can expedite this time-consuming yet crucial task. This study demonstrates an end-to-end damage detection framework that has two components: a semi-automatic machine to crush the canola seeds to prepare the samples rapidly for detecting the damages and a supporting object detection algorithm based on a compressed YOLOv5s model. The lightweight detection model was deployed in an Edge-AI device and integrated with the crusher to detect the damages in real-time. Two Convolutional Neural Network (CNN) architectures, viz. ShuffleNetV2 and MobileNetV3 were compared in terms of model parameters, computational cost, and model size to replace the standard CSPDarknet53 CNN backbone present in YOLOv5s. The best-performing model was deployed into an NVIDIA Jetson Nano embedded device for real-time inferencing. Results indicate that in terms of model size, the ShuffleNetV2_YOLOv5s model was 56.5 and 80.4 % smaller than the MobileNetV3_YOLOv5s and the baseline YOLOv5s models, respectively. While, in terms of computational cost, the ShuffleNetV2_YOLOv5s model was 64.1 % and 99.5 % less expensive than the MobileNetV3_YOLOv5s and the baseline YOLOv5s models, respectively. This study demonstrated an end-to-end framework for detecting damages in canola supported by a real-time and lightweight damage detection model.