Shield tail clearance images feature extraction based on semi-supervised semantic segmentation

Abstract Manual extraction of shield tail clearance image features is usually subjective, time-consuming and limited. Deep learning methods have shown tremendous potential in shield tail clearance images feature extraction, and can overcome the shortcomings of manual feature extraction effectively. However, most of the current deep learning methods need a huge volume of labelled data, which constrains their widespread use in practical engineering. In this paper, we propose a semi-supervised learning model-Semi-SegSTC to achieve feature extraction for shield tail clearance image. The method consists of a student model and a teacher model, which have the same structure as well as weight initialization. The two models are based on DeepLabV3+ network, which firstly adopts ResNet-50 as the backbone network to improve the segmentation accuracy while reducing the network parameters, and secondly integrates the criss-cross attention mechanism to improve the network's ability to extract the shield tail clearance image features by increasing the weights of the effective channels in the image. Firstly we input the weakly enhanced unlabelled image into the model for prediction to obtain pseudo-labels, retaining the pseudo-labels only when the model produces a high confidence prediction, and using the prediction obtained by inputting this pseudo-labels to the model after supervising a strongly enhanced version of the same image. We conduct extensive experiments on the shield tail clearance images dataset as well as the PASCAL VOC 2012 dataset, and the results show that our proposed method consistently exhibits excellent performance when trained with limited labelled data.