基于Retinex-Net的集气站指针仪表暗光图像增强方案

Low-light image enhancement and reading of pointer meter in gas gathering station based on Retinex-Net

集气站在引入巡检机器人进行仪表检测与识别任务时,对图像质量及处理实时性要求较高,但实际工作过程中因光照、遮挡等因素,经常拍摄到难以识别的暗光图像,传统暗光增强方法难以同时满足增强效果与实时性需求。基于Retinex理论及神经网络,文中引入Retinex-Net,在集气站实地拍摄的正常光/暗光图像集上进行测试,实现对暗光仪表图像的增强,在恢复仪表的真实色彩的同时尽可能的保留指针等细节信息。将该算法与CLAHE、MSRCR、AutoMSRCR算法在处理效果与处理速度上进行了对比,分析了Retinex-Net用于暗光仪表图像增强的优越性;对比处理前后仪表图像读数情况,证明了算法的实用性与鲁棒性。

Instrument detection and identification in gas gathering station require real-time processing and high-quality images when inspection robot is introduced.Because factors like illumination and occlusion,the robot captures low-light images frequently in actual working process.Traditional methods could not meet the requirements of enhancement effect and real-time performance at the same time.Based on the Retinex theory and neural network,we applied Retinex-Net to normal/low-light image set taken on the station,and realized the enhancement of the low-light images,which restored the true color of the instrument as same as remained details such as the pointer.Comparing our algorithm with CLAHE,MSRCR and AutoMSRCR in processing effect and speed,we analyzed the superiority of Retinex-Net.In addition,the practicability and robustness of the algorithm are proved by comparing the readings of meters before and after processing.