Parallel Concatenated Feature Pyramid Network for Dehazing a Single Image on Smartphone Images

Abstract

Smartphones capturing images in outdoor environments are often affected by adverse weather conditions, resulting in low-quality images. This paper introduces the Parallel Concatenated Feature Pyramid Network (C-FPN) to address the challenge of dehazing single smartphone images. Dehazing a single image on smartphones is considered an ill-posed problem. While the Feature Pyramid Network (FPN) is widely used in computer vision tasks, its feature extraction is limited by the max-pooling operator. Furthermore, it cannot retain the hazy feature and restore the image at the same time, which fails to preserve critical hazy image features. Additionally, most existing methods struggle to balance preserving haze-relevant information with effective image restoration. To address these limitations, this study proposes a novel parallel concatenated FP architecture that estimates atmosphere light and calculates transmission information on smartphones. The key contributions of this paper include (1) designing a parallel concatenated FP architecture capable of retrieving hazy features across various environments in deeper layers, (2) incorporating a concatenation structure to retain hazy information, enabling depth estimation and the generation of a transmission map, (3) using the transmission map as an input for a convolutional neural network with a dehazing loss function to calculate atmosphere light under different environments, and (4) implementing a skipping connection in the C-FPN to retain essential features, facilitating an end-to-end learning structure. The proposed method demonstrates superior performance on the SOTS, NH-HAZE 2, and synthetic hazy image indoor datasets. The PSNR/SSIM achieve 26.58/0.948, 26.28/0.966 and 17.15/0.761, respectively. In addition to dehazing, the method achieves excellent object detection performance.