CNN with Pose Segmentation for Suspicious Object Detection in MMW Security Images.

Area of Science:

• Imaging Science
• Computer Vision
• Security Technology

Background:

• Millimeter-wave (MMW) imaging offers non-contact security screening but faces challenges in detecting small, diverse suspicious objects.
• Conventional methods using Convolutional Neural Networks (CNNs) struggle with limited datasets and object recognition complexities.

Purpose of the Study:

• To develop a more effective algorithm for suspicious object detection in MMW images.
• To address the limitations of conventional object recognition approaches in MMW security screening.

Main Methods:

• A new algorithm integrating human pose segmentation with deep CNN detection was developed.
• The method leverages human body segmentation to train CNNs, focusing on detecting abnormal patterns indicative of suspicious objects.
• Suspicious object recognition was reframed as a binary classification task.

Main Results:

• The proposed algorithm demonstrated optimal performance in detecting suspicious objects within MMW images.
• The approach achieved high effectiveness and superiority compared to conventional methods.
• Experiments confirmed the algorithm's ability to identify abnormal patterns associated with concealed items.

Conclusions:

• The CNN-based algorithm with pose segmentation offers a concise and high-performing solution for MMW suspicious object detection.
• This method enhances security screening by improving the accuracy and efficiency of identifying concealed threats.