Awesome

Occluded Prohibited Items Detection: An X-ray Security Inspection Benchmark and De-occlusion Attention Module

This repository is the official implementation of De-occlusion Attention Module(DOAM) and OPIXray dataset. Paper link: https://arxiv.org/abs/2004.08656.

DOAM framework:

DOAM

OPIXray dataset:

Overview
Dataset
Install
Usage
Results
Contributing
Acknowledgement
License

Overview

Security inspection often deals with a piece of baggage or suitcase where objects are heavily overlapped with each other, resulting in an unsatisfactory performance for prohibited items detection in X-ray images.

In this work, first, we contribute a high-quality dataset named OPIXray, each of which is annotated manually by professional inspectors from an international airport. To the best of our knowledge, this is the first dataset specifically designed for object detection in security inspection. Second, we propose a De-occlusion Attention Module (DOAM) that can be inserted as a plug-and-play module into most detectors, aiming at detecting occluded prohibited items in X-ray images. Finally, we evaluate our method on the OPIXray dataset and compare it to several baselines, including popular methods for detection and attention mechanisms. As is shown from the results, our proposed method significantly outperforms existing models.

Dataset

In this project, we build the first dataset specially designed for occluded prohibited items detection in security inspection.

All images of OPIXray dataset are scanned by security inspection machine and annotated manually by professional inspectors from an international airport, and the standard of annotating is based on the standard of training security inspectors.
OPIXray dataset contains a total of 8885 X-ray images(7019 for training, 1776 for testing), including 5 categories of cutters, namely, Folding Knife, Straight Knife, Scissor, Utility Knife, Multi-tool Knife.
<div align=center> <img src="images/statistics.png" width="50%" height="50%"/> </div>
In order to study the impact brought by object occlusion levels, we divide the testing set into three subsets and name them Occlusion Level 1 (OL1), Occlusion Level 2 (OL2) and Occlusion Level 3 (OL3), where the number indicates occlusion level of prohibited items in images.
<div align=center> <img src="images/occlusion_distribution.png" width="50%" height="50%"/> </div>
The information structure of annotation file is as follows: image name, category, top-left position of prohibited item (x1, y1), bottom-right position of prohibited item (x2, y2).

Install

Install pytorch

The code is tested on python3.5 and pytorch1.3.1

Clone the repository

git clone https://github.com/OPIXray-author/OPIXray.git

Access OPIXray dataset

OPIXray dataset is available only for ACADEMIC PURPOSE, if you are interested in our dataset, feel free to contact rstao@buaa.edu.cn.

Usage

Training

change root to DOAM
```
cd DOAM
```

modify OPIXray_ROOT in data/OPIXray.py :

OPIXray_ROOT = 'OPIXray_Dataset/train/'	# path to trainset of OPIXray Dataset

run train.py
```
python train.py --image_sets 'OPIXray_Dataset/train/train_knife.txt' --transfer './weights/ssd300_mAP_77.43_v2.pth' --save_folder './checkpoints'
```
- --save_folder: the path to save the checkpoint while training.
- --dataset_root: the path to the training dataset.
- --image_sets: the path to a TXT file that saves all the picture names used for training.
- --transfer: indicates the path to SSD pre-trained weight(available at here).

Testing

change root to DOAM
```
cd DOAM
```

modify OPIXray_ROOT in data/OPIXray.py :

OPIXray_ROOT = "OPIXray_Dataset/test/"	# path to testset of OPIXray Dataset

run test.py
```
python test.py --imagesetfile 'OPIXray_Dataset/test/test_knife.txt' --trained_model './weights/DOAM.pth' 
```
- --imagesetfile: the path to a TXT file that saves all the picture names used for testing.
  - test_knife.txt: total testset
  - test_knife-1.txt: occlusion level 1 (no or slight occlusion).
  - test_knife-2.txt: occlusion level 2 (partial occlusion).
  - test_knife-3.txt: occlusion level 3 (severe or full occlusion).
- --trained_model: the checkpoint you want to evaluate, and our trained model is available at here (password: m9zk).

Results

The performance of SSD and YOLOv3 under three different object occlusion levels:

The amount changes of performance improvement of DOAM with occlusion level increases:

Performance comparison between DOAM-integrated network and baselines for three famous detection approaches:

Contributing

Acknowledgement

Thank amdegroot/ssd.pytorch for the SSD backbone on which our DOAM based.

License

The images and the corresponding annotations in OPIXray Dataset can be used ONLY for academic purposes, NOT for commercial purposes.