Image Dehazing and Deraining

Haze and rain damage images in different ways. Atmospheric scattering adds veiling light and reduces contrast with distance. Rain can create streaks, splashes, droplets or blur, and a wet lens may cover scene information entirely. Dehazing algorithms estimate how much scene radiance has been attenuated and how much atmospheric light was added. The dark channel prior research by He, Sun and Tang is an influential single-image approach based on statistics of haze-free outdoor images. Its assumptions can fail on bright or unusual scenes, so an output is an estimate rather than a direct measurement of hidden colors. Deraining methods target streaks or other rain patterns. The CVPR work on deep joint rain detection and removal is one research example for single images, including heavy accumulation in its tested setting. A model can mistake thin scene structures for rain and erase them. Conversely, strong rain or droplets can hide detail that no single frame contains. Video offers temporal information, but moving cameras and objects complicate alignment. Do not generalize success on simulated rain streaks to every windshield, night scene or fog condition. Evaluation should be task-specific. A pleasing dehazed landscape can still have distorted colors or amplified noise. An object detector may improve on some scenes and worsen on others. Compare outputs with available clean references or repeated real-world captures, inspect small structures, and report failure cases by weather intensity and lighting. If the processed image informs driving or safety decisions, test the complete perception stack and provide a fallback when visibility is too poor. Weather removal should preserve the original image and processing record. It can help a human see existing evidence but cannot certify a license plate or person hidden behind an opaque drop. Communicate uncertainty rather than smoothing away an occlusion and presenting invented content as fact.