A new image-sequence haze removal system based on DM6446 Davinci processor

Abstract

Poor visibility due to haze poses a challenge for driving and can significantly compromise safety. In this paper, we present an algorithm and an implementation that assists the driver by providing an electronic view that improves visibility via haze removal. Our optimized implementation on the Texas Instruments TMS320DM6446EVM DSP based evaluation board provides full D1 video resolution and frame processed for haze removal every 1.7 seconds. This near real-time implementation can significantly augment the drivers vision by providing updates at a rate comparable with the rate at which the driver looks at the rear-view mirror. The algorithm uses a dark channel prior with an edge-guided filter for fast implementation; both components are specifically modified for improving performance on the DM6446 processor. The system provides the first DSP-based embedded implementation of haze removal processing with near real-time performance and represents a significant speed-up and reduction in memory requirements over previously reported algorithms.