US 12277682 B1Patent grantUnited States

Haze-Adaptive Vision for Intelligent Vehicles

Official patent title

Method for image dehazing of vehicular images

Arabic title: طريقة لإزالة الضباب من صور المركبات

Invention

Problem

Fog, clouds, rain, smoke, pollution, and other airborne particles reduce image contrast, alter colors, and hide details. Many dehazing datasets and models address only one haze type, limiting reliability for vehicle and computer-vision systems operating across changing conditions.

Why it matters

Clearer images can support perception tasks such as object detection, segmentation, remote sensing, surveillance, and autonomous navigation when atmospheric visibility is degraded.

Approach

The method creates multiple synthetic haze datasets, trains several haze-type classifiers and specialized dehazers, classifies an image from an autonomous-vehicle camera using a hybrid conditional classifier, and routes the image to the dehazer specialized for the predicted haze class.

Who may benefit

Potential beneficiaries include autonomous-vehicle and advanced-driver-assistance developers, traffic and security surveillance operators, remote-sensing teams, robotics companies, and computer-vision platform providers.

Potential value

Instead of applying one general dehazer, the architecture first identifies the atmospheric condition and selects a specialized model. It also supports multiple classifiers and configurable haze types, intensities, datasets, and target domains.

Background

Atmospheric haze reduces visibility by scattering transmitted light, degrading contrast, color, and fine detail. This affects both human interpretation and machine-vision tasks in vehicles, satellites, surveillance systems, and mobile platforms. Existing datasets commonly represent only fog, cloud, or an unspecified single haze type, while many algorithms are trained as universal dehazers. Such systems may not account for condition-specific image degradation. The disclosure addresses this limitation by separating haze-type classification from specialized restoration and by training on scene groups with multiple synthetic atmospheric effects.

Technology overview

Ground-truth images are split into outdoor, street, farmland, and satellite sets and overlaid with cloud, fog, environmental haze, rain, and, for selected sets, smoke at different intensities. Multiple classifiers and cloud-, fog-, haze-, smoke-, and rain-specific dehazers are trained. For an autonomous-vehicle image, a hybrid conditional classifier combines predicted classes and probabilities through condition blocks to determine the haze class. The corresponding specialized network then removes the detected condition from the image.

Potential applications

Image enhancement for autonomous vehicles and driver-assistance systems.
Traffic, perimeter, and security-camera visibility improvement.
Remote-sensing and satellite-image restoration.
Robotic and mobile-platform computer vision in adverse weather.
Preprocessing for object detection and image segmentation.

Evidence-supported advantages

Classifies haze type before selecting a restoration model.
Uses specialized dehazers for cloud, fog, environmental haze, smoke, and rain.
Combines multiple classifier outputs through a hybrid conditional framework.
Supports configurable scene domains, haze types, and intensity levels.
Targets images acquired directly from autonomous-vehicle cameras.

Development stage

Dataset generation, classifier and dehazer architecture, training workflow, and visual comparisons are described; quantitative validation on diverse real vehicle data and deployment hardware was not established in the supplied evidence. The development stage was not independently verified.

Commercial opportunity

The architecture may be licensed to automotive-perception, surveillance, remote-sensing, or edge-vision providers. Productization requires real-world multi-weather datasets, quantitative safety-relevant benchmarks, latency and power evaluation on vehicle hardware, failure detection, integration with downstream perception, cybersecurity controls, and field validation across locations and cameras.

Patent classifications

WIPO IPC

G06T5/60Image data processing or generation, in general
G06T5/73Image data processing or generation, in general

CPC

G06T5/73Image data processing or generation, in general
G06V20/50Image or video recognition or understanding
G06T5/60Image data processing or generation, in general
G06V10/764Image or video recognition or understanding
G06V10/87Image or video recognition or understanding
G06V10/774Image or video recognition or understanding
G06V10/945Image or video recognition or understanding
G06T2207/30232Image data processing or generation, in general
G06T2207/30252Image data processing or generation, in general
G06T2207/10032Image data processing or generation, in general
G06T2207/30181Image data processing or generation, in general
G06T2207/20081Image data processing or generation, in general
G06T2207/20092Image data processing or generation, in general
G06T2207/20084Image data processing or generation, in general

Inventors

First inventorMd Tanvir Islam
InventorIk Hyun Lee
InventorAbdul Khader Jilani Saudagar
InventorAbdullah Altameem
InventorMohammed Abaoud
InventorKhan Muhammad

Keywords

image dehazing
autonomous vehicle
haze classification
hybrid conditional classifier
deep learning
computer vision
adverse weather
specialized dehazer

Patent document and drawings

The patent publication is mapped to this record. Patent drawings remain within that publication; no separately cleared public media package has been supplied.