Augmentation Matters: A Mix-Paste Method for X-Ray Prohibited Item Detection under Noisy Annotations

TL;DR

This paper addresses the challenging problem of training a robust prohibited item detector under noisy annotations from a novel perspective of data augmentation, and proposes an effective label-aware mixed patch paste augmentation method (Mix-Paste), which demonstrates the great potential of data augmentation to handle noise annotations.

Abstract

Automatic X-ray prohibited item detection is vital for public safety. Existing deep learning-based methods all assume that the annotations of training X-ray images are correct. However, obtaining correct annotations is extremely hard if not impossible for large-scale X-ray images, where item overlapping is ubiquitous.As a result, X-ray images are easily contaminated with noisy annotations, leading to performance deterioration of existing methods.In this paper, we address the challenging problem of training a robust prohibited item detector under noisy annotations (including both category noise and bounding box noise) from a novel perspective of data augmentation, and propose an effective label-aware mixed patch paste augmentation method (Mix-Paste). Specifically, for each item patch, we mix several item patches with the same category label from different images and replace the original patch in the image with the mixed patch. In this way, the probability of containing the correct prohibited item within the generated image is increased. Meanwhile, the mixing process mimics item overlapping, enabling the model to learn the characteristics of X-ray images. Moreover, we design an item-based large-loss suppression (LLS) strategy to suppress the large losses corresponding to potentially positive predictions of additional items due to the mixing operation. We show the superiority of our method on X-ray datasets under noisy annotations. In addition, we evaluate our method on the noisy MS-COCO dataset to showcase its generalization ability. These results clearly indicate the great potential of data augmentation to handle noise annotations. The source code is released at https://github.com/wscds/Mix-Paste.

Figures (6)

• Figure 1: Examples in an X-ray dataset wei2020occluded. (a) Examples with noisy category labels (the correct labels from left to right are folding knife, utility knife, and utility knife, respectively). (b) Examples with inaccurate bounding boxes. (c) Examples with correct annotations in X-ray images, where item overlapping is ubiquitous.
• Figure 2: Overview of our proposed method for training a robust prohibited item detector under noisy annotations. (a) illustrates Mix-Paste which mixes multiple item patches with the same category label (the correct label in the upper item patch is the scissor) for data augmentation. (b) illustrates the LLS strategy which suppresses the large losses corresponding to potentially positive predictions of additional items during loss calculation. $\mathcal{PB}_{neg}$: the predicted bounding boxes whose IoUs between them and the ground-truth bounding boxes are less than a threshold; $\mathcal{PB}_{fb}$: the predicted bounding boxes whose IoUs are greater than a threshold and the predicted label is the background; $\mathcal{PB}_{pos}$: the predicted bounding boxes whose IoUs are greater than a threshold and the predicted label is the same as the ground-truth category label; $\mathcal{PB}_{pp}$: the predicted bounding boxes whose IoUs are greater than a threshold and the predicted label (not the background) is different from the ground-truth category label.
• Figure 3: (a) Ablation study results on the influence of the number of patches for Mix-Paste on the OPIXray dataset. (b) Training curve comparison between our Mix-Paste and the baseline method on the OPIXray dataset.
• Figure 4: Examples of generated images on the OPIXray dataset. (a) The generated images with the linear combination. (b) The generated images with the edge smoothing mask.
• Figure 5: Performance comparison between Mix-Paste and the competing methods at different noise rates under (a) category noise and (b) bounding box noise on the OPIXray dataset.