A CNN-based lightweight ensemble model for detecting defective carrots

Carrot grading plays a crucial role in producing high-competitive carrot products. However, carrot grading mainly depends on manual work nowadays, and this unreliable situation leads to high labor consumption, low efficiency, and unstable standard. In this study, a lightweight model (CarrotNet) based on machine vision with DCNN was proposed with the inspiration of several classic CNNs. The optimizations were conducted for the influential hyper-parameters in CarrotNet through comparative analysis. And the partial layers of the network were removed by the ablation study to obtain a more efficient structure. Furthermore, ensemble learning was adopted in the model to further improve the model accuracy. In the test set, the proposed model CarrotNet gained an accuracy of 97.04%, the modeling time of 1.42 h, the model size of 8.18 MB, and the detection speed of about 80 images per second. The robust performance of CarrotNet in the carrot dataset indicates that it can be used in on-line detection and grading of carrot external quality. (c) 2021 IAgrE. Published by Elsevier Ltd. All rights reserved.