Misty gas leak detection was considered critical in fields such as chemical engineering and aviation. To address the problem of detecting leaks in visible-light images when misty gas leak data are scarce and specific features are lacking, a visual detection method based on multiple features was developed. Differences in image characteristics before and after a misty gas leak were analyzed, and pixel-level features including transmittance, coherence, and uniformity were designed to characterize the leak. The decision mechanism of support vector machine (SVM) was combined with exponential smoothing; confidence optimization and smoothing factor constraints were incorporated to weight the ensemble algorithm, forming an adaptive classification model for misty gas leaks. This model accurately determines misty gas leak states in visible-light images. Validations in typical leak scenarios show that this method effectively identifies gas leak status, achieving a success rate above 95% with an average processing time of around 35 ms per frame, meeting real-time detection requirements.