A comparative analysis of DETR and Conditional DETR for identifying smoking scenes

Smoking remains a significant global public health concern, leading to over 8 million deaths annually and significantly increasing the risk of various diseases. Despite the effectiveness of policies like banning smoking scenes in media and implementing public smoking bans in controlling smoking behaviors, the proliferation of digital media presents new challenges in managing exposure to smoking-related content. In response to these challenges, this study introduces an automated solution that leverages advanced computer vision techniques, specifically utilizing DETR and Conditional DETR models with ResNet-50 backbones, to detect and apply Gaussian blur to smoking scenes in videos. The performance of these models was rigorously evaluated using mean Average Precision (mAP) metrics across datasets containing 4,000, 3,000, and 2,000 images. The results demonstrated that the DETR model achieved the highest AP of 58.5 with the 4000-image dataset, surpassing the Conditional DETR model, which achieved an AP of 55.0 with the 3000-image dataset. Furthermore, in video processing tasks, the DETR model exhibited superior efficiency, with an average processing time of 56.8 seconds, compared to Conditional DETR's 59.2 seconds. The implementation of Gaussian blur effectively obscured smoking activities in the video content, preserving the overall context while significantly reducing the visibility of harmful content. These findings underscore the advantages of the DETR model in terms of both detection speed and accuracy, particularly when applied to larger datasets. This study highlights the potential of DETR in enhancing automated content moderation tools, contributing to the ongoing efforts to mitigate the public health risks associated with smoking in the digital age.