Dynamic System-Based Image Enhancement and Segmentation for Fire Scene Analysis

Abstract

To address the limitations of traditional fire scene image processing systems, such as insufficient edge detection accuracy and poor denoising performance, this paper proposes a comprehensive image processing framework based on dynamic models. The core contributions include: an image enhancement model is built based on a nonlinear dynamic diffusion process, which introduces a dynamic adjustment term to achieve adaptive denoising while preserving edges. An image segmentation model that improves the Geodesic Active Contour Model is built by constructing a gradient-adaptive extended geodesic activity contour model, significantly enhancing its capability to handle weak edges and complex structures. Experiments were conducted on specialized fire image datasets. For enhancement, the Fire Scene Image Enhancement (RFSIE) dataset and the NTIRE20 dataset were used. For segmentation, the FLAME dataset and the Fire Segmentation Dataset were employed. The proposed enhancement model improved the Peak Signal-to-Noise Ratio (PSNR) by approximately 20.0% compared with median filtering on the RFSIE dataset and by 25.0% on the NTIRE20 dataset. The segmentation model achieved an accuracy of 5.6% higher than that of the Fully Convolutional Network (FCN) on the FLAME dataset. Furthermore, in image classification tasks, the proposed model improved the accuracy of flame, smoke, and background classification by about 30.8%, 16.1%, and 8.6%, respectively, compared with median filtering. The research demonstrates that the dynamic system provides a more efficient and robust solution for fire scene analysis, with significant potential for application in fire monitoring and rescue operations.