Division of datasets.

There are few data labels in the agricultural field, and accurate annotation of existing data requires professional knowledge and is time-consuming and laborious, especially the images collected in the actual greenhouse scene, which lack accurate annotation by professionals. Fine-grained refers to the highly detailed division or analysis of data or tasks, with a particular emphasis on capturing micro-level differences. In order to improve the accuracy of greenhouse crop disease identification, the crop disease identification problem is regarded as a fine-grained classification problem, and the attention mechanism is introduced into the classification network. The VAE enhancement strategy is introduced into the disease identification network model to improve the accuracy when the annotation is insufficient. Aiming at the problem that the actual environmental background of greenhouse is complex, there are many disturbances, the disease spot area is small, and the difference between leaf disease and wilt and soil is not obvious, a fine-grained identification model of leaf disease based on reconstruction-generation is further proposed. The attention mechanism was used to increase the recognition ability. During training, the VAE strategy was first used to make full use of a large number of labeled and unlabeled data to realize unsupervised learning, and then the labeled data was used for supervised disease identification, and the Reconstruction-Generation Network(RGN) was used to force the classification network to pay more attention to discriminative regions to find differences. Reconstruction-generation belongs to self-supervised learning, which uses the unsupervised information in the data to construct supervised signals, and can generate useful feature representations by learning the structure and pattern in the data. Experimental results show that the classification recognition accuracy of the proposed fine-grained leaf disease identification model based on reconstruction-generation adopts the attention mechanism reached 98.03%. The proposed method is applied to the detection model, the correct recognition rate of diseased leaves was 95.07%, and the correct recognition rate of healthy leaves was 98.46%, which can realize the end-to-end detection and identification of diseased leaves and meet the practical requirements.