Zero-shot image dehazing method based on high-frequency enhancement and depth-guided optimization

Abstract

The application discloses a zero sample image defogging method based on high-frequency enhancement and depth guidance optimization, first extracts high-frequency components from a single foggy image by using Laplacian pyramid decomposition, and inputs the foggy image and the high-frequency components into a defogging network, dynamically weights and fuses high-frequency detailed information and low-frequency structural information through a multi-scale feature adaptive enhancement module, estimates the transmittance and the atmospheric light parameter based on an estimation network, and generates a defogging image. Further, a re-fogging network symmetrical to the defogging network structure is introduced to generate a re-fogging image, a bidirectional mapping relationship between defogging and fogging is constructed, and the network is jointly optimized through cycle consistency, atmospheric light consistency, transmittance consistency and depth consistency constraint based on difference perception. The method effectively improves the detail fidelity and structural consistency of the defogging result without paired supervised samples, and is especially suitable for image enhancement tasks in complex non-uniform fog and haze scenes.

Images