A few-shot segmentation method for prohibited item inspection.

Background: With the rapid development of deep learning, several neural network models have been proposed for automatic segmentation of prohibited items. These methods usually based on a substantial amount of labelled training data. However, for some prohibited items of rarely appearing, it is difficult to obtain enough labelled samples. Furthermore, the category of prohibited items varies in different scenarios and security levels, and new items may appear from time to time.

Objective: In order to predict prohibited items with only a few annotated samples and inspect prohibited items of new categories without the requirement of retraining, we introduce an Attention-Based Graph Matching Network.

Methods: This model applies a few-shot semantic segmentation network to address the issue of prohibited item inspection. First, a pair of graphs are modelled between a query image and several support images. Then, after the pair of graphs are entered into two Graph Attention Units with similarity weights and equal weights, the attentive matching results will be obtained. According to the matching results, the prohibited items can be segmented from the query image.

Results: Experiment results and comparison using the Xray-PI dataset and SIXray dataset show that our model outperforms several other state-of-the-art learning models.

Conclusions: This study demonstrates that the similarity loss function and the space restriction module proposed by our model can effectively remove noise and supplement spatial information, which makes the segmentation of the prohibited items on X-ray images more accurate.