Study On Jujube Classification Method Based On Visual Attention Mechanism

Xinjiang is located in the northwest,with sparsely populated areas.The ‘hometown of melons and fruits’ has long been famous.Jujube has become a large-scale crop in Xinjiang because of its nutritional value,economic value and ecological value of wind prevention and sand fixation.In recent years,the jujube industry has developed rapidly and the output has been continuously innovating.However,the classification methods of jujube market are different and the mixed grade is serious,which not only misleads consumers ’ independent choice,but also affects the process of commodity circulation.At present,manual and mechanical methods are commonly used in grading,but there are problems such as high labor intensity,high production cost,single and unstable grading standard,easy to cause secondary damage of fruit,and insufficient defect detection.In recent years,with the rapid development of computer vision technology,the application field of deep learning has been expanding,which provides a new idea and method for jujube grading.This paper takes Alal Grey Jujube as the research object to further explore the classification method of jujube based on machine vision,and proposes the visual saliency fusion method for defect jujube detection.Combined with the characteristics of jujube wrinkle,size and color,the comprehensive classification of defect-free jujube is realized.The research in this paper is mainly reflected in the following aspects :(1)Clear grading standards and establish jujube data set.By referring to the national and industrial standards of jujube classification,the classification standard of grey jujube was further clarified,.The jujube image acquisition device was built,and the number of samples was expanded by data enhancement.The Otsu segmentation method was used to remove the image background.The image pixel value was transformed into [ 0,1 ] by normalization operation.The labelme tool was used to manually label the data,and the experimental data set was established.(2)A method for defect detection of jujube based on explicit integration was proposed.In order to more accurately determine the types of defects,a fusion detection method based on convolution neural network was proposed.The detection method was composed of global detection network and local detection network.These two branches were based on the migration learning of convolution neural network pre-trained by big data.The Res Net50 network with cavity convolution was used for global saliency extraction,and the Vgg19 network of Dropout layer was added for local feature extraction.Then,the least square method was used for the fusion of local and global saliency,and the final saliency map of defect jujube was obtained.Finally,the defect jujube was classified.(3)A multi-attention mixed jujube grading method was proposed.Under the convolution attention mechanism,the interaction among channel attention,spatial attention and channel-spatial attention modules is explained,and the multiple attention mixing module is proposed.Based on Dense Net121,the multi-attention mixing module was introduced to construct three branch networks,and the spatial attention map,the channel attention map and the channel-spatial attention map were generated respectively.The attention maps of the three branch networks were added to calculate the average value and output the final attention map for jujube grading.Through confusion matrix analysis,combined with network training loss function and classification accuracy curve,the classification accuracy of the method is 95.68 %.