Deconvolution guided semi-supervised plant leaf disease identification and segmentation method

Abstract

The invention provides a deconvolution-guided semi-supervised plant leaf disease identification and segmentation method, which uses a small amount of disease category labels and disease spot pixel-level labels to achieve disease category identification and disease spot region segmentation through deconvolution. According to the method, a category prediction label of an unmarked sample is generatedthrough a consistency regularization and entropy minimization method; image mixing is carried out on the marked sample and the unmarked sample, and semi-supervised disease classification is carried out by utilizing the newly generated image; and up-sampling is performed on the category information, and semi-supervised scab segmentation is performed by using a small number of pixel-level marks. Inthe process of model training, model parameters are updated by using exponential weighted average, so that the model is more robust in test data. The method is suitable for identifying and segmentingplant leaf diseases with insufficient label samples, integration of identification and segmentation is achieved, the model has high generalization capacity in leaf images with insufficient light andforeign matter shielding, and the identification and segmentation speed can meet the real-time requirement.

Description

technical field [0001] The invention belongs to the field of plant disease detection, in particular to a semi-supervised plant leaf disease identification and segmentation method guided by deconvolution. Background technique [0002] Diseases are one of the main reasons affecting the growth of crops. Timely analysis of the characteristics of crop disease spots can help to quickly give corresponding disease prevention guidance and suggestions to eliminate disease alarms. At present, there are mainly two types of methods for classification of plant diseases. One is to use the artificially designed plant disease feature extraction method, and use the machine learning method to classify the extracted features. This method generally needs to segment the diseased spots or diseased leaves first, which increases the workload in the early stage, and it is necessary to target each time. Different disease combination design feature extraction methods lack robustness, and it is difficu...

AI Technical Summary

Problems solved by technology

One is to use artificially designed plant disease feature extraction methods, and use machine learning methods to classify the extracted features. This method generally needs to segment disease spots or leaves first, which increases the workload in the early stage, and needs to be targeted every time. Design feature extraction methods for different disease combinations, lack of robustness, and it is difficult to distinguish similar diseases

The second method is to use deep convolutional neural networks to automatically extract disease features for classification, but these end-to-end classifiers require a large amount of labeled data, and the feature learning process of convolutional neural networks is opaque and easy to overfit. Learning useless features will interfere with disease identification

For plant disease segmentation, currently the semantic segmentation network is mainly used to divide the image into background and lesion areas. Although the semantic segmentation network can segment different types of diseases at the same time, it has two shortcomings: first, the semantic segmentation network requires a large number of pixels Second, the lesions of various diseases are similar. When doing the segmentation of lesions, they can only be divided into two types: background and lesions. It is difficult to determine the type of disease to which the lesions belong.

For convolutional neural networks, the feature extraction process of traditional models cannot be manually intervened. When overfitting occurs, the model will pay too much attention to irrelevant features.

Images