Abstract
To develop a high-performance convolutional neural network (CNN) model for plant image classification automatically, we propose a neural architecture search (NAS) method tailored to multifunctional cells (MFC), termed MFC-NAS. Initially, a search space based on MFC is designed, encompassing transfer cell, normal cell, pooling cell, and dropout cell, with transfer cell dedicated to exploring weight-sharing layers. Subsequently, an MFC-oriented search strategy is adopted: different shallow blocks from pre-trained models such as MobileNet V3 are searched to construct transfer cell. Similar strategies are applied to pooling cell, dropout cell, and normal cell, exploring diverse pooling types and sizes for pooling cell and various dropout rates for dropout cell. Finally, the best-found cells are stacked to form a plant image classification CNN based on MFC. Experiments conducted on two publicly available plant image datasets demonstrate that MFC-NAS achieves the optimal cells after approximately 69 GPU-hours of search. Compared to state-of-the-art (SOTA) methods like ResNet-50 and EfficientNet, this approach attains higher accuracy (~ 99.10%) with an average single-sample inference time of around 12.6 ms. Moreover, the number of network parameters used in the proposed method is only 6.9% of ResNet-50's (approximately 1.58 M).