Semester

Fall

Date of Graduation

2024

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Jeremy Dawson

Committee Member

Brian Powell

Committee Member

Prashnna Gyawali

Abstract

Many cargo containers enter the United States every day by truck, rail, and sea. As a result of the large number of cargo containers entering the United States, not all of them can be thoroughly inspected. Most of these containers contain properly documented and legal cargo, but some people take advantage of this situation by hiding illicit items in the cargo containers such as drugs. To more efficiently and thoroughly inspect cargo containers, Customs and Border Protection (CBP) uses X-ray imaging machines to obtain images that reveal the interior of cargo containers. These X-ray images must be inspected to ensure that there are no illicit items in the cargo containers. An algorithm that inspects these images will make the detection of illicit items more reliable and more efficient. This work proposes a self-supervised, knowledge distillation, patch-based algorithm to address this problem. When knowledge distillation is applied to anomaly detection, there is a teacher network that generates an output and a student network that aims to replicate the output of the teacher network for normal images. The proposed method has two teachers, one pretrained and one adapted for the target dataset, one student, and a segmentation network. The teachers and student are used to create multi-level anomaly maps and the segmentation network is used to combine these anomaly maps. Additionally, pseudo anomalies are used to create a self-supervision task for the model. This model was evaluated using the CargoX [30] dataset, which is a dataset that contains X-ray images of cargo containers with synthetic anomalies.

Download

Share

COinS