Elevating Perception: Blobs-Based Dehazing with Image Enhancement Using Mask-Guided Fully Convolutional Networks and Multi-Scale Label Smoothing
Zhiqiang Wang, Aryan Joshi, Tianfu Guo, Wenjia Ren, Kaining Zhang · 2024
This paper presents a framework for image dehazing and enhancement that combines blobs-based model generation, YOLO for object detection, and mask-guided convolutional networks (CNNs) with multi-scale label smoothing. By leveraging these components, our approach aims to improve dehazing performance and enhance image clarity. The blobs-based detection focuses on haze-affected regions, while YOLO enables rapid object localization. The integration of mask-guided CNNs with multi-scale label smoothing further refines the output, resulting in superior visual quality and detail preservation. Our early results suggest a speed increase of about 5 to 10% compared to traditional methods. User feedback indicates that our retouching techniques perform better than established methods, even when using unpaired training data.