Deep tobacco leaf grading using ShuffleNet and self-adaptive attention

This work proposes a self-adaptive attention (SAA) module that improves the image classification ability of the convolutional neural networks (CNNs) in tobacco leaf grading. The method achieves automated grading of tobacco leaves according to the Chinese national standard. The SAA module employs micro-factorized pointwise and depthwise convolutions, respectively, to compute the channel-based weights and make a balance between channel and node connectivity. It achieves least redundancy between input and output and focuses on the essential features that affect the grading performance. Experimental results validate that the grading accuracy of different CNNs can all be improved by the SAA module. The best grading accuracy 90.91% is achieved when combining the SAA module with ShuffleNetV1, which is considered to be satisfactory in objective tobacco leaf grading.