An Image Restoration Method for Improving Matching Robustness of Indoor Smoke Scene

Abstract

Smoggy interference caused by indoor fires makes machine vision technology challenging to apply in the fire rescue field. Smoke and condensed water vapor aerosol from suppression activities limit visibility, making image matching difficult. To overcome this problem, an image restoration method for indoor smoke scenes is proposed. First, the dark channel prior algorithm for indoor smoke scenes is improved, and the atmospheric light estimation method is optimized by combining the density peak clustering algorithm and position constraint. A model update approach is also advanced to achieve real-time dehazing of image sequences. Afterward, the effect of photometric changes caused by the image restoration on matching is analyzed. The feature matching is performed using the pyramid Lucas–Kanade (LK) optical flow method, while the random sampling consistency algorithm is used to eliminate outliers. Finally, an indoor smoke dataset is created to evaluate the algorithm, and a comprehensive analysis of the algorithm's limitations is conducted to provide a thorough understanding of the algorithm's potential shortcomings. The evaluations confirm that the proposed method can effectively improve the robustness and accuracy of indoor smoke scene image matching. The percentage increase in robustness is close to 100%, and the accuracy has increased by 10%. Overall, this approach holds practical value for the fire rescue field, and it may encounter limitations in handling scenarios with dense smoke, dark smog, and dynamic flames. Further improvements and optimizations are required to address these challenges.