A Novel CNN Based Dual-Stream Network with DeepLabv3+ for Accurate Fire and Animal Detection in Foggy Environments

Fire and animal detection under extremely aggravated atmospheric fog conditions is very challenging due to low visibility, noise, and environmental hindrances. These create significant obstacles in the reliability and accuracy of detection systems, which are vital for wildlife conservation and disaster management. This study proposes a Dual-Stream Network with DeepLabv3+ to enhance certainty and robustness under poor foggy conditions. The framework includes Multi-Scale Fusion Dehazing Network (MSFDNet) for dehazing, Non-Local Means (NLM) for denoising, and data augmentation techniques. DeepLabv3+ with enhanced Atrous Spatial Pyramid Pooling (ASPP) ensures precise segmentation. Feature extraction uses Spider Wasp Optimization (SWO) and Hippopotamus Optimization Algorithm (HOA). Detection integrates YOLOv8-SE, Enhanced FocalNet, and Convolutional Neural Networks-Long Short-Term Memory-Multi-Head Self-Attention (CNN-LSTM-MSA) for reliable classification performance. The proposed system achieves a notable improvement, reaching 98.16% accuracy, demonstrating its effectiveness for detection under severe fog conditions.