Depth-Wise and Depth-Wise Separable YOLO Models for Concealed Object Detection Using Terahertz Images

Abstract

Terahertz imaging is highly effective for detecting concealed objects due to its non-harming nature and its ability to penetrate materials like clothes, paper and plastic, etc. In environment where detection technologies are limited, terahertz imaging emerges as one of the effective and safest methods available. Unlike techniques such as X-rays, it does not emit harmful radiation, making it suitable for surveillance applications. However, many existing object detection models are computationally intensive which can hinder their deployment in real time or resource constrained environments. To address this issue, traditional convolutional operations in the deep learning models have been replaced with depth-wise convolutions and depth-wise separable convolutions in proposed approach. These modifications significantly reduce the number of trainable parameters and computational load during model training. The optimized architecture has been integrated into widely used object detection models – namely YOLOv5m and YOLOv8m, using terahertz images of concealed objects as input. This integration enhances training efficiency with minimal loss in accuracy, making the models more suitable for deployment on devices with limited computational power and memory.