DDSNet: A dual-channel dual-mode smoke detection network for lightweight industrial scenarios

In recent years, vision technology has been increasingly utilized in environmental monitoring, particularly in air quality prediction and smoke detection. Despite advancements, there remains a need for efficient models capable of accurate real-time detection on resource-constrained devices. This paper presents dual-channel dual-mode smoke detection network (DDSNet), a novel dual-channel architecture designed for detecting industrial smoke from single images. DDSNet leverages high-frequency and low-frequency channels to capture detailed and broad-spectrum features, respectively. Extensive experiments on the D-fire and Fire-Flame-Dataset demonstrate that DDSNet achieves better performance with an accuracy of 98.30% and 99.49% respectively, while maintaining low inference time and parameters. Our ablation study confirms the effectiveness of the dual-channel structure in enhancing detection accuracy without significantly increasing computational complexity, making DDSNet suitable for deployment on mobile devices.