Deep learning for X-ray image analysis

• Main Author: Ren, Bing
• Other Authors: Wen Bihan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/141477
• Institution: Nanyang Technological University

Deep learning is a branch of machine learning, which is an efficient way to achieve the goal, artificial intelligence. It takes the idea of human neural networks and learns the features of a large dataset, which contributes greatly to language and image analysis. This dissertation applied deep learning for X-Ray image analysis is to utilize deep learning methods to implement object detection. Security checking for bags and luggage at airports and the real-way station is done by the staff, which is inefficient and requires heavy human workload. The dissertation aims to solve this problem. The outcome of the dissertation is a system that achieves automated detection of prohibited items. It aims to locate the prohibited items and identifies their categories. Raw X-ray images of security checking are already prepared by researches, i.e., the large-scale X-ray images dataset for security inspection which is publicly available. It contains 6 categories of prohibited objects. To achieve position detection some pre-processing (labeling) is required to allocate the position and classification of objects. The dissertation gave a review of different deep learning methods for object detection, like R-CNN, YOLO, and SSD. With labeled X-Ray images, the experiments used SSD and YOLO respectively to implement prohibited item detection and made a comparison of their task performances, such as accuracy and time consumed. As the exiting dataset of the annotation is limited, it could result in overfitting. The dissertation tries to figure out some methods for improvement, like reducing the network complexity, data augment, and early stopping. When using the exciting label datasets provide by other experts, the SSD network and YOLOv3 can achieve 0.1683 and 0.1457 mAP. The experiments were carried on ii making some improvements to the performance of SSD. The existing dataset contains some inaccurate labels and mistaken label. Another dataset was prepared which is labeled manually with careful selection of labels and the position of bounding boxes. It helped to raise the mAP of SSD to 0.2156. Later, the further experiments were conducted tried to improve the class imbalance by manipulating datasets, which further improved the mAP of SSD by 0.0109. Besides, the experiments also utilized a resized smaller feature extraction network by fixing a certain convolutional layer of VGG16, which can be interpreted as reducing the network complexity. The project used the existing dataset prepared by researchers to train the resized SSD network, and the result approved that, for this particular X-ray project, this method did not have much effect on the mAP of SSD models. Moreover, the dissertation provided some perspective for further development. In the real scenario, only a small percentage of luggage contains prohibited items, which is regarded as a positive sample. Those carrying no prohibited items are negative samples. Thus, it is necessary to take the ratio of positive samples and negative samples into consideration. To further improve the performance of the object detection model, a dataset can be constructed with a certain proportion of positive and negative samples, that mimics the real scenario of prohibited item detection at security checkpoints.