Features Representation for Flue-cured Tobacco Grading Based on Transfer Learning to Hard Sample

In the study of an automatic tobacco-grading system based on machine vision, features extraction of tobacco images is a critical step. The more effective features extracted, the better classification accuracy got. Most existing works have shown that apparent features have an effect on tobacco grading, which include features of color, shape and texture. However, tobacco posture usually is unstable in the process of collecting tobacco images, which may lead to these features information unreliable. In this work, a brief framework of tobacco images classification is proposed. To extract the better features, we fuse apparent features and low-dimensional deep features. When learning deep features, the modified AlexNet network is used to transfer the shallower convolutional features. And we fine-tune the network with hard sample mining actively. Experiments show that fusing apparent features and deep features compared to just apparent or deep features can improve the accuracy of tobacco-grading. In addition, compared with adding samples randomly, hard-mining actively can achieve the same performance with fewer samples.