Task-driven Image Preprocessing Algorithm Evaluation Strategy

The image preprocessing algorithm can effectively improve performance of the advanced visual algorithms under complex conditions. But the evaluation of the image preprocessing algorithm usually uses the traditional PSNR/SSIM indicators, there is a big gap between the evaluation result and the performance of the algorithm in actual vision tasks. This article uses the UE4 virtual engine and referring to the atmospheric scattering model to construct a haze vehicle small object dataset (VSOD). Based on virtual data, a vision task-driven image preprocessing algorithm evaluation strategy is proposed. We take dehazing algorithms and object detection tasks as examples to verify the task-driven evaluation strategy. By analyzing the evaluation results of task-driven evaluation strategy and traditional PSNR/SSIM, we proved that the task-driven evaluation strategy can more accurately evaluate the performance of image preprocessing algorithms in actual vision tasks.