In modern agriculture, the detection of maize cobs is highly accurate and efficient for yield estimation, crop management and resource allocation. It is labor intensive, and less than ideal for in processing of automation. In order to deal with such challenges, we present this work on MaizeNet, a high speed detection system based on lightweight CNNs adapted to work well in resource constrained environments and on edge devices. To obtain good performance yet low computational efficiency, we design a custom CNN architecture for MaizeNet. The dataset is diverse and contains thoroughly annotated ground truth data and is used to train and validate the model. In maize cob detection, MaizeNet achieves a mean average precision (mAP) of 91.4% and processes 25 FPS on standard mobile hardware in real time. It is shown through comparative evaluations that MaizeNet outperforms all other deep learning models, and through an ablation study of architectural choice on model performance, the value of certain architectural choices on model performance. The robustness under the diverse field conditions and scalability to the larger agricultural datasets brought out the utility of MaizeNet as a useful tool for precision agriculture. MaizeNet improves agricultural productivity and sustainability by contributing to improving accuracy and efficiency in management of a maize cropping system.
Mishra et al. (Fri,) studied this question.