Research On Flue-cured Tobacco Classification Algorithm Based On Deep Learning

In the process of production and acquisition of flue-cured tobacco,tobacco quality sorting is currently artificially finished by the relative workers.The strong subjectivity and low grading efficiency of workers have given rise to the research on automatic sorting technology for tobacco leaves.With the development of machine vision,tobacco classification based on vision algorithm has become a fast-growing research field.However,there are still some problems in existing researches.The data obtained from the existing online image acquisition system are limited and they are different from the ideal data.In addition,the classification accuracy of traditional methods with manual features is relatively low.Deep features extracted by deep learning technology are effective to improve the performance of classification,but a large number of labeled data are needed for training,and it is difficult to obtain extensive labeled tobacco images.In response of these problems,pre-processing steps of tobacco images are designed and deep convolutional neural network and transfer learning are employed for tobacco leaves grouping and grading.The main contributions are as follows:An image preprocessing method for tobacco is proposed by analyzing all tobacco images collected on the site.In preprocessing,median filtering is used to reduce image noise.Through the calculating of the b-channel frequency histogram of "Lab" space,the foreground part of tobacco image is effectively segmented by histogram bimodal method.In order to delete anomaly images and get more ideal training data,tobacco leaf postures are analyzed by morphological method.In this paper,tobacco leaf grouping and grading models are proposed based on deep convolutional network.In response to the problem of less training data and single training categories,the method of two-step modifying network parameters based on Alexnet model is proposed to prevent over-fitting.In our method,local response normalization layer is deleted and batch normalization layer is added to optimize the network structure.In order to further optimize the number of labeled samples used in the training process,the method of learning actively and incrementally is proposed,with which,hard samples are mined by calculating the information entropy so redundant samples are reduced in training.Comparison experiments on central orange tobacco leaves show that the parameter optimization could make the model fit better and the network structure optimization could improve the Loss oscillation problem on training.That is,the training optimization actively and incrementally can train a better model with less data.Tobacco leaves grading algorithm based on transfer learning is proposed to optimize the performance of grading.Because of less training data and imprecise tobacco leaves grading results,we fine-tune pre-trained model of tobacco leaves grouping to express low-dimension deep features of tobacco.Then,combing them with color,shape and texture features based on traditional method to represent tobacco characterizes for grading task.Considering that the collection of tobacco image is real-time and different batch images affect the accuracy of classification,the online transfer learning algorithm is proposed.An old classifier is learned with the existing training data,and a new classifier is learned for the new data.When a new image is received,its label is got by weighting of prediction results for two classifiers.Moreover,weights of the two classifiers are updated online based on the case of the wrong samples.Experiments show that the feature extraction method based on fusion transfer learning and traditional features is better than the method with only deep features or traditional features;the online transfer learning algorithm is better than machine learning algorithm with only new training data,and the error rate of algorithm is smaller.