Contextual Style Coherence Network for X-Ray Prohibited Item Image Synthesis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Technology

Background:

• Effective prohibited item detection in X-ray baggage images is vital for security.
• Current methods struggle with data scarcity, requiring costly manual annotation.
• Existing image synthesis techniques like Threat Image Projection (TIP) lack contextual style coherence, reducing realism.

Purpose of the Study:

• To develop a novel method for synthesizing realistic X-ray prohibited item images.
• To address the limitations of existing TIP methods regarding style coherence and sample diversity.
• To improve the performance of prohibited item detection systems through enhanced training data.

Main Methods:

• Proposed a Contextual Style Coherence Network for X-ray image synthesis.
• Introduced a style fusion module for foreground-background coherence.
• Utilized feature-space TIP, affine transformation for diversity, and Gram matrix for style normalization.
• Employed mask partial convolution for seamless inpainting and boundary transition.
• Implemented an unsupervised adversarial training pipeline with adversarial loss and total variation regularization.

Main Results:

• Generated synthetic X-ray images with improved quality and diversity.
• Demonstrated enhanced style coherence between foreground items and background images.
• Achieved better object detection performance on prohibited item detection datasets compared to existing methods.
• Validated the effectiveness of the proposed style fusion and inpainting techniques.

Conclusions:

• The Contextual Style Coherence Network effectively synthesizes realistic X-ray prohibited item images.
• The proposed method overcomes data insufficiency limitations in training security inspection systems.
• This approach significantly improves the performance of prohibited item detection, contributing to enhanced security.