Research And Implementation Of Silkworm Disease Recognition System Based On Convolutional Neural Network

Silkworm breeding is an important economic source in Guangxi,China.However,silkworm disease has a huge impact on it.If the silkworm disease can not be identified and diagnosed timely and accurately,it will cause direct economic losses to silkworm farmers and affect the stable development of social economy in Guangxi,China.With the progress of image processing technology,the maturity of machine learning theory and the popularization of mobile Internet life,it is possible to realize the intelligent and scientific recognition of silkworm disease in Guangxi region by identifying silkworm diseases images by Convolutional Neural Network algorithm and combining with mobile application.However,this work is challenged by the difference of image hierarchical information which is caused by the different distance of silkworm disease image acquisition,and the abundance of detailed features in the spot area of silkworm body disease.In order to obtain the characteristic information of silkworm disease image better and realize the accurate recognition of silkworm disease image,this paper proposes an attention-based image retrieval algorithm for silkworm disease.And on this basis,combined with the research of the project of automatic recognition system of silkworm diseases in Guangxi Zhuang Autonomous Region Sericulture Technology Promotion Terminal,a silkworm disease recognition system based on Android platform was designed and implemented,which has been put into trial operation for half a year with good testing results.The main research work completed in this paper is as follows:(1)Based on the characteristics of the great difference of hierarchical information and the abundance of detailed features in silkworm disease image,this paper proposes a concatenation dense network based on attention mechanism(Attention Concatenation Dense Net,AC-Dense Net)by improving the model structure.By setting the contrast experiments,the retrieval results of different Convolutional neural network algorithms are compared and analyzed,and the recognition rate of AC-Dense Net was 85.6%,which was improved significantly.(2)To solve the problem of the number of samples is unbalanced and some classes difficult to be recognized in dataset of silkworm diseases,based on the Focal loss function,combined with the multiple classification problems of silkworm diseases,using improved multi focal loss function.By setting the contrast experiment,the recognition results of AC-Dense Net under different loss functions were compared and analyzed,the recognition rate of the improved algorithm was further improved to89.3%.(3)The recognition of silkworm disease in the past mainly relied on the experience of silkworm disease experts,which was inefficient and limited in time and space,combined with the research on image recognition algorithm of silkworm diseases,based on Andriod platform and Flask framework,a C/S based silkworm disease automatic recognition system was designed and implemented.The recognition rate of the system reached 87.8% by using the test set of 600 silkworm disease images,and the average retrieval time was 2.4s which met users needs well.In summary,the algorithm for image recognition of silkworm disease has achieved good results in recognition rate,and the recognition system of silkworm disease has reached the expected target in recognition accuracy and efficiency,which proves the reliability and practicability of this recognition system.