DTDG-Net: A Few-shot Data Augmentation Method for X-ray Security Inspection Images

Since the detection of prohibited items is a low-probability event, the target detection model based on deep learning often suffers from issues of low recognition accuracy due to limited training data. To ad-dress this challenge, this article proposes an enhanced Domain Transformation Data Generation Network (DTDG-Net), aiming to solve the problems brought by the scarcity of training samples. Initially, we designed a Color Consistency Cycle Generative Adversarial Network (CC-CycleGAN) based on a Dynamic Color Attention Module (DCAM), which can effectively ensure the color consistency between the original and generated images. Secondly, we developed a Data Fusion Generator (DFG), which aligns the transparency features of the generated and original images through a blending Poisson fusion method based on hue transformation, ultimately achieving data augmentation with limited samples. Finally, to objectively evaluate the effectiveness of our approach, we introduce a style transfer quality evaluation co-efficient, specifically designed for the performance evaluation of few-shot sample data augmentation for object detection. Through this metric, we can comprehensively and accurately measure the contribution of data augmentation to the object detection task. This unique framework design holds potential for widespread applications in the field of prohibited item detection in X-ray images. Through a series of experiments, we demonstrate that our approach, exhibits stronger robustness and generalization, effectively alleviating the challenges posed by few-shot sample problems, providing an innovative solution for pro-hibited item detection in X-ray images.