Tobacco mildew detection is a crucial aspect of the tobacco industry in China. Detection techniques based on semi-supervised semantic segmentation have attracted significant attention due to their low cost and high efficiency. Therefore, this study designed a semi-supervised segmentation model tailored for complex tobacco mildew detection scenarios, and conducted research from both the semi-supervised strategy and model perspectives. In terms of the semi-supervised strategy design, this study employed the standard teacher–student framework, focusing on data perturbation to achieve satisfactory results without introducing additional complexity and cost. First, the CutMix technique was improved to ensure high accuracy in predicting weakly perturbed images and to implement strong perturbation strategies more effectively. Second, a simple feature perturbation method was proposed as a supplement to image perturbation to further explore the perturbation space. In terms of the segmentation model design, this study enhanced the widely used DeepLabV3+ in the semi-supervised domain. To address the issue of losing fine structural information, the down-sampling convolution in the encoder was replaced with dilated convolution, and a multi-scale decoder was introduced to fully utilize the multi-stage features of the encoder, ensuring a comprehensive understanding of the images. Extensive experiments conducted on a self-made dataset demonstrated the effectiveness and superiority of this method in tobacco mildew detection.
Xu et al. (Mon,) studied this question.