Identification of tomato leaf diseases based on multi-channel automatic orientation recurrent attention network

Abstract

In existing deep learning-based tomato leaf disease identification algorithms, there are two factors limit the performance: 1) noise can be easily generated during image acquisition, transmission, and processing, which makes it challenging to extract disease features; 2) inter-class similarity and intra-class variability of tomato leaf diseases make it challenging to identify disease images. A new identification model for tomato leaf disease is proposed in this paper to solve the above problems. First, an Asymptotic Non-Local Means algorithm (ANLM) is introduced to reduce the image's noise interference and to decrease the difficulty of extracting tomato leaf disease features in the identification network. Then, a Multi-channel Automatic Orientation Recurrent Attention Network (M−AORANet) is proposed to extract abundant disease features. An automatic orientation attention network is designed to locate lesion sites on tomato leaves. The fine multiscale feature is extracted and recycled to solve the problem of inter-class similarity and intra-class variability identification of tomato leaf diseases. Experimental results on 7493 images demonstrated that the identification accuracy of M−AORANet reached 96.47%, which outperformed other current identification networks in comparison experiments. It can effectively provide decision information for tomato disease identification systems in precision agriculture.