CASIA-Net: An indoor work site smoking detection framework

Abstract

Detecting smoking behaviors in indoor work sites poses significant challenges due to the small scale of targets, poor visibility, and cluttered environments. These factors significantly heighten the risk of fire hazards. We propose the Context-Aware Small-Item Attention Net (CASIA-Net), a novel detection framework to address these issues. CASIA-Net incorporates a Deformable Feature Extraction (DFE) module to tackle the non-salient characteristics of smoking targets. It adaptively adjusts the convolution kernel size according to the target scale. An Adaptive Feature Attention (AFA) module is proposed to extract small objects. It enhances the attention to critical features of smoking from complex indoor work site backgrounds. To address the issue of attention drift in complex environments, a Smoker Feature Integration module is proposed to integrate the features extracted by DFE and AFA. Additionally, a dedicated dataset for indoor work site smoking detection is constructed. Experimental results demonstrate that the proposed model achieves an mAP50 of 0.917 on the dataset with a compact weight of 6MB. The proposed model demonstrates outstanding accuracy, robustness, and lightweight design. It is highly suitable for deployment in complex indoor work sites and industrial applications.