LSANNet: A lightweight convolutional neural network for maize leaf disease identification

Abstract

Maize (Zea Mays) is a major food crop and is of great importance to ensure national food security. However, maize leaf diseases occur from time to time, which poses a serious threat to grain yield and quality, so methods for the quick identification of maize leaf diseases are particularly important. In this paper, a long-short attention neural network (LSANNet) is proposed for maize leaf diseases identification. The main component of the LSANNet is the long-short attention block (LSAB). The long-short connection method enables the fusion of multi-scale features, which enhances the model generalisation capability. The attention mechanism is applied in the block, which aims to enhance the extraction of maize leaf features. The effectiveness of separable convolution and attention modules is demonstrated by ablation studies. Experimental results on 124 unseen images show that the accuracy of the proposed model on the test sets reaches 94.35%, which is better than the accuracy of existing models, such as VGG16, ResNet50, DenseNet201, MobileNetV3S, and Xception. The practical performance of the proposed network model is verified by deploying the model on a mobile device, demonstrating strong compatibility and high recognition. In this paper, a lightweight convolutional neural work is proposed for maize leaf disease identification, and the performance of the network on the test sets meets the required requirements. This research will provide an idea for the identification of maize leaf diseases and disease prevention schemes for agricultural production.