Application Of Deep Learning Driven Computer Vision Method In Crop Growth Situation And Disease Diagnosis

Crop growth status assessment and disease detection are necessary links in agricultural production and an important part of realizing smart agriculture and agricultural automation.The traditional method of disease detection and growth stage judgment mainly relies on the experience of farmers and agricultural experts.This manual method is not only inefficient and timeconsuming,but also difficult to promote.With the development of deep learning,computer vision driven by deep learning has been widely used in agriculture due to its fast,accurate,and nondestructive advantages.In this paper,a deep learning-driven computer vision method is applied to the assessment of paddy growth status and the detection of apple disease with the conditions of insufficient and unbalanced training samples.The main research contents of this paper are as follows:（1）Based on the channel attention and the spatial attention,a three-channel Siamese neural network TC-SNet was proposed to detect apple diseases with the condition of small samples.The effectiveness of the TC-SNet network structure was verified by comparing the small-sample performance with multiple baseline models.In addition,the effects of four basic image processingbased data augmentation methods on the performance of the TC-SNet model were also discussed,verifying that the model had strong robustness.The comprehensive evaluation results of TC-SNet showed that with the condition of 25 training samples per class,the accuracy of the network was as high as 98.5%.Finally,the TC-SNet was applied to the real environment in the wild to detect apple diseases.The results showed that TC-SNet also had excellent few-shot learning ability in complex environments.（2）A generative adversarial network Attention GAN was proposed based on the multi-head attention and the spatial attention,and improved the plant disease detection performance with the condition of unbalanced disease data.By comparing the quality,diversity,and similarity between the generated images and the original images of the Attention GAN and the baseline model DCGAN,the rationality and superiority of the Attention GAN network structure were proved.The results showed that when Attention GAN was applied to the task of plant disease detection and classification,the accuracy of disease detection was improved to 96.81%.With the condition of unbalanced disease data,using Attention GAN to expand the data of the unbalanced class,the accuracy of the unbalanced class was increased from 81.82% to 95.56%,which effectively solved the problem of low accuracy of the unbalanced class.（3）Based on computer vision feature extraction method and support vector machine,a nondestructive detection method for the complete growth stage of paddy and the corresponding evaluation index GSR were proposed.Firstly,the RGB image of paddy was converted to HSV color space,and the distribution of pixel frequency in H channel with different growth stages of paddy was discussed and analyzed.On this basis,threshold segmentation was performed on the green pixels and yellow pixels in the image to obtain the green area index and the yellow area index,which were used as color features.Combined with the images of different paddy growth stages,this experiment analyzed the rationality of the extracted color features and texture features,and verified that the texture features extracted from paddy images based on gray level cooccurrence matrix could further improve the accuracy of paddy growth situation assessment.The results showed that GSR could accurately evaluate the growth stage of paddy on farmland,and its detection accuracy was 98%.