Abstract:

Crop diseases and insect pests cause a large amount of economic losses in agricultural production activities, and it is difficult to meet field production requirements of the current society if only relying on manual surveys by agronomist staffs. Applications of the machine vision technology can realize the automatic classification and detection of crop diseases and insect pests, and provide the guarantee for accurate and efficient agricultural productions. However, existing detection methods based on the deep learning framework and convolutional neural networks are constrained by factors such as rigid convolutional receptive field, inefficient data enhancement operator, and small sample size. In order to make up for the above shortcomings of existing detection technologies in term of recognition accuracy, a method for the classification of agricultural economic crop diseases and insect pests based on the masked autoencoding learning paradigm is proposed in this study. Through local random content masking, semantic feature extraction, and global context reconstruction of high-dimensional mapping of input crop images, the proposed algorithm can fully mine implicit representations of high-level semantics of images and model the long-distance contextual relationship in the same map, so as to train a more robust model with less data samples. Moreover, the model eliminates the interference of the high-frequency noise on the pre-training feature extraction processing by the relative total variational transformation. The results of comparison between the proposed method and current methods based on mainstream convolutional networks show that the proposed method can significantly improve the performance of existing methods, and the accuracy rate is improved from 90.48% to 95.24% based on ResNet50 benchmark network.

Key words: machine vision, deep learning, agricultural cash crop, diseases and pests detection, masked autoencoding, relative total variation, neural network, convolutional receptive field