Improved Model for Smoke Detection Based on Concentration Features using YOLOv7tiny

Abstract

Smoke is often present in the early stages of a fire. Detecting low smoke concentration and small targets during these early stages can be challenging. This paper proposes an improved smoke detection algorithm that leverages the characteristics of smoke concentration using YOLOv7tiny. The improved algorithm consists of the following components: 1) utilizing the dark channel prior theory to extract smoke concentration characteristics and using the synthesized αRGB image as an input feature to enhance the features of sparse smoke; 2) designing a light-BiFPN multi-scale feature fusion structure to improve the detection performance of small target smoke; 3) using depth separable convolution to replace the original standard convolution and reduce the model parameter quantity. Experimental results on a self-made dataset show that the improved algorithm performs better in detecting sparse smoke and small target smoke, with mAP@0.5 and Recall reaching 94.03% and 95.62% respectively, and the detection FPS increasing to 118.78 frames/s. Moreover, the model parameter quantity decreases to 4.97M. The improved algorithm demonstrates superior performance in the detection of sparse and small smoke in the early stages of a fire.