In today’s world, agricultural products are becoming increasingly scarce globally due to a variety of factors, and the early and accurate automatic identification of plant diseases can help ensure the stability and sustainability of agricultural production, improve the quality and safety of agricultural products, and help promote agricultural modernization and sustainable development. For this purpose, a lightweight deep isotropic convolutional neural network model, FoldNet, is designed for plant disease identification in this study. The model improves the architecture of residual neural networks by first folding the chain of the same blocks and then connecting these blocks with jump connections of different distances. Such a design allows the neural network to explore a larger receptive domain, enhancing its multiscale representation capability, increasing the direct propagation of information throughout the network, and improving the performance of the neural network. The FoldNet model achieved a recognition accuracy of 99.84% on the laboratory dataset PlantVillage using only 685k parameters and a recognition accuracy of 90.49% on the realistic scene dataset FGVC8 using only 516k parameters, which is competitive with other state-of-the-art models. In addition, as far as we know, our model is the first model that has fewer than 1M parameters while achieving state-of-the-art accuracy in plant disease identification. This proposal facilitates precision agriculture applications on mobile, low-end terminals.
Although neural-network architectures are critical for their performance, how the structural characteristics of a neural network affect its performance has still not been fully explored. Here, we map architectures of neural networks to directed acyclic graphs (DAGs), and find that incoherence, a structural characteristic to measure the order of DAGs, is a good indicator for the performance of corresponding neural networks. Therefore, we propose a deep isotropic neural-network architecture by folding a chain of the same blocks and then connecting the blocks with skip connections at different distances. Our model, named FoldNet, has two distinguishing features compared with traditional residual neural networks. First, the distances between block pairs connected by skip connections increase from always equal to one to specially selected different values, which lead to more incoherent graphs and let the neural network explore larger receptive fields and, thus, enhance its multi-scale representation ability. Second, the number of direct paths increases from one to multiple, which leads to a larger proportion of shorter paths and, thus, improves the direct propagation of information throughout the entire network. Image-classification results on CIFAR-10 and Tiny ImageNet benchmarks suggested that our new network architecture performs better than traditional residual neural networks. FoldNet with 25.4M parameters can achieve 72.67% top-1 accuracy on the Tiny ImageNet after 100 epochs, which is competitive compared with the-state-of-art results on the Tiny ImageNet.
Although neural network architectures are critical for their performance, how the structural characteristics of a neural network affect its performance has still not been fully explored. We here map architectures of neural network to directed acyclic graphs, and find that incoherence, a structural characteristic to measure the order of directed acyclic graphs, is a good indicator for the performance of corresponding neural networks. Therefore we propose a deep isotropic neural network architecture by folding a chain of same blocks then connecting the blocks with skip connections at different distances. Our models, named FoldNet, have two distinguishing features compared with traditional residual neural netowrks. First, the distances between block pairs connected by skip connections increase from always equal to one to specially selected different values, which lead to more incoherent graphs and let the neural network explore larger receptive fields and thus enhance its multi-scale representation ability. Second, the number of direct paths increases from one to multiple, which leads to a larger proportion of shorter paths and thus improve the direct propagation of information throughout the entire network. Image classification results on CIFAR-10 and Tiny ImageNet benchmarks suggested that our new network architecture performs better than traditional residual neural networks.