LMNet: A lightweight MobileNetV2-based model for identification of rice leaf diseases

Various diseases have greatly affected the yield and quality of rice, and the pathogenic microbes for most of diseases infect and damage the leaves. It is an urgent requirement to accurately and timely monitor and diagnose the severities for precise spraying. As a non-destructive and non-contact technology, the image-based analysis technology has been widely applied to the identification, detection and classification of rice leaf diseases. The emergence of convolutional neural network (CNN) has facilitated the improvement of efficiency and accuracy, especially in recent years. Nevertheless, the existing models usually show the characteristics of high complexity, many parameters involved and high computational power requirements, which restrict the recognition efficiency of mobile phones, laptops and other mobile terminals. It is necessary to timely detect and accurately identify the types and severities of diseases in rice fields. Therefore, lightweight CNNs and universal backbone networks are needed to be studied thoroughly. A recognition and classification method with low computing power for rice diseases was studied, which can be transplanted to mobile devices.In this study, a lightweight convolutional neural network (CNN) model named Lightweight MobileNetV2-based (LMNet) model was designed. It can identify and classify rice leaf diseases, such as rice blast and flax leaf spot, based on the images taken by mobile devices in the field. LMNet is constructed by using convolutional block and inverted residual module. According to different convolution strides of inverted residual modules, the ECA (Efficient Channel Attention) attention modules are added to the LMNet and multi-scale feature fusion structures are designed. Compared with other attention mechanisms, ECA is extremely simple in thought and operation and has minimal impact on network processing speed. Combined with channel attention, the model focuses on important features and reduces the impact of complex environment. In addition, a multi-scale feature fusion module is designed to change the size of convolution kernel, and fuse the feature graph extracted from the $3 \times 3$ inverted residual module with the convolution kernel of $5 \times 5$ inverted residual module with the feature graph extracted from the linear $1 \times 1$ pointwise convolution to reduce the feature loss caused in the process of up- and down-sampling operations.The experimental results show that the Accuracy of proposedLMNet model on the test data set reaches 98.24%, which is better than classical CNN models of VGG16, ResNet50 and lightweight CNN models of MobileNetV1 and MobileNetV2. In addition, the model size of LMNet is only 2.33 MB, which is only 6% of the MobileNetV1 model. It is much easier to be embedded into the portable handheld mobile devices. The study can provide a methodological reference for identifying other crop leaf diseases via a lightweight CNN model. It provides multi-type training samples and model parameter tuning for macro-scale monitoring.