Automatic Identification Research Of Tobacco Diseases Based On Convolutional Neural Network

Tobacco disease is one of the biggest problems of tobacco growers, and our country is a big country of tobacco production so that it is an important source of tax revenue. Tobacco diseases do not control and not prevention, will seriously affect the income of farmers and the country's fiscal and tax. According to statistics, China had planted 1610 acres of flue-cured tobacco in 2015, industrial and commercial of the tobacco industry taxes reached 1.14 trillion Yuan, the annual of total 1 trillion Yuan turned over to finance. The quality of tobacco leaves depends on whether tobacco healthy growth. But in the process of the growth of tobacco, inevitably affected by various factors, including all kinds of the dangers of tobacco diseases, once produce disease, will let the quality of tobacco leaves of different level, affect the overall level of tobacco leaf. Traditional tobacco disease recognition need professional and technical personnel and technical equipment, unable to meet the needs of farmers, not satisfy the large area promotion.In this paper, a variety of tobacco diseases was collected under field environment. Interactive segmentation algorithm was used to accurately extract the tobacco image spots. The shoot equipment is android system mobile phone. In this paper, using the structure of the convolutional neural network method, combined with back propagation algorithm on tobacco disease image recognition classification. Convolution neural network is the fusion of the technique of artificial neural network and deep learning and it can implement local perception, hierarchical, feature extraction and recognition classification characteristics of combining the global training. This article uses the convolutional neural network model is 6 layer structure, including the input and output layer, two layer convolution layers, two layer sampling layers and one connect fully layer. The structure by means of local wild, sharing and drop weight sampling method to identify the image displacement deformation and distortion. Finally, this article through the relevant experiment verifies the feasibility of convolution neural network model and high performance. At the same time, we discussed the different number of iterations and the influence of different resolution of network model training and classification.This paper designed a tobacco diseases recognition based on Web is designed. When found tobacco diseases in the field users took pictures and uploaded images to server. Server automatically recognized the type of tobacco diseases and fed back to the user's mobile clients including recognition result and control methods. Users based on the feedback results to guide the production of tobacco.