Abstract:The ExposureNet project addresses the issue of image blur caused by imbalanced camera exposure settings, by developing an autonomous system for controlling these settings. The system, trained comprehensively, predicts ideal exposure based on the semantic features of a scene, using only shutter speed and ISO as training signals. This approach leads to faster performance, reduced computational costs, and lower power consumption, with experimental results showing significant improvements over existing methods. The quality of images we perceive visually is heavily impacted by the settings used for camera exposure. When these settings are imbalanced, it can result in an undesired prominent phenomenon known as blur effects. To address this problem, an ExposureNet project has been undertaken, which aims to develop an autonomous camera exposure settings control system for blur effects prevention. The proposed ExposureNet model is a CNN/Transformer hybrid neural structure, created and trained in a comprehensive manner to effectively predict the ideal exposure settings based on the semantic features of the scene being captured. This system is designed to learn the necessary steps for processing, such as identifying relevant scene features, using only two camera exposure parameters (shutter speed (SHS) and ISO) as training signals. As a result, this system can associate the semantic features of a scene with the appropriate exposure parameter adjustments, customized to the scene's dynamics and lighting conditions. By simultaneously optimizing all processing steps and bypassing traditional post‐processing stages, the proposed system is designed to achieve faster performance, reduced computational cost, and lower power consumption. Experimental results demonstrate that the proposed system significantly outperforms existing methods and achieves cutting‐edge performance.

Neural Camera Simulators

A Lensless Camera Simulator via Deep Learning

All-day Thin-Lens Computational Imaging with Scene-Specific Learning Recovery

LightSim: Neural Lighting Simulation for Urban Scenes

SpikingSIM: A Bio-Inspired Spiking Simulator.

Non-Uniform Exposure Imaging via Neuromorphic Shutter Control

An End-to-End Real-World Camera Imaging Pipeline

Deep Learning for Camera Data Acquisition, Control, and Image Estimation

A Portable Multiscopic Camera for Novel View and Time Synthesis in Dynamic Scenes

A system for generating complex physically accurate sensor images for automotive applications

A Physics-based Noise Formation Model for Extreme Low-light Raw Denoising

OASim: an Open and Adaptive Simulator based on Neural Rendering for Autonomous Driving

Inverting the Imaging Process by Learning an Implicit Camera Model

Physics-Guided ISO-Dependent Sensor Noise Modeling for Extreme Low-Light Photography

Event Camera Simulator Design for Modeling Attention-based Inference Architectures

ExposureNet: Mobile camera exposure parameters autonomous control for blur effect prevention

NeuroZoom: Denoising and Super Resolving Neuromorphic Events and Spikes.

Simulation-driven Learning: a Deep Learning Approach for Image Scanning Microscopy Via Physical Imaging Models

Neural Real-Time Recalibration for Infrared Multi-Camera Systems

Soft Prototyping Camera Designs for Car Detection Based on a Convolutional Neural Network

Physics-based Noise Modeling for Extreme Low-light Photography