考虑低光照场景的自适应路面病害检测模型

An Adaptive Pavement Defect-detection Model Considering Low-light Scenarios

为提升低光照环境下路面病害识别模型的精确度与鲁棒性,提出了一种考虑低光照场景的自适应路面病害检测模型。将采集的数据按照亮度分成低光照图像和正常光照图像2类,随后对低光照图像进行自适应增强,并接入路面病害检测流程。为解决现有光照增强图像失真的问题,在Zero-DCE++基准模型上引入光照约束网络,提出了光照增强模块Zero-DCE-Retinex。在此基础上,以提升路面病害检测精度为核心需求,提出了以病害识别模型为驱动的光照增强模块训练策略。具体的,在训练过程中,将病害识别模型嵌入光照增强模块,并通过损失函数来引导光照增强模块的训练。试验结果表明:当使用所提模型对低光照图像进行增强后,再经过YOLOv7模型进行检测,比直接输入低光照图像的方式,AP指标提升了7.04%,F1-score指标提升了11.22%;图像质量亦得到进一步改善,表现在图像信息熵、标准差和平均梯度的提升分别为0.62、13.24和21.61。与同类无监督光照模型相比,所提模型在目标检测任务的AP和F1-score提升方面显著优于其他模型,同时有效避免了增强结果出现色彩失真和过度曝光等问题。因此,提出的模型可广泛适用于复杂道路光照场景,拓展了现有病害检测模型的应用环境,具有重要的工程应用价值。

This study proposed an adaptive pavement defect-detection model to enhance the accuracy and robustness of pavement defect-identification models in low-light conditions.Data were collected and categorized into low-light and normal-light images based on brightness.Adaptive enhancement was applied to low-light images,which were integrated into the pavement defect detection process.To address the issue of distortion in existing illumination-enhanced methods,we introduced an illumination-constrained network onto the Zero-DCE++baseline model,resulting in a luminance-enhancement module called Zero-DCE-Retinex.Building upon this,a training strategy for the luminance-enhancement module driven by the defect-recognition model was proposed to improve pavement defect-detection accuracy.During training,the defect recognition model was embedded into the illumination-enhancement module,guiding its training through a loss function.The proposed model was used to enhance low-light images and subsequently detect them using the YOLOv7 model,achieving an average precision increase of 7.04%,and an increase of 11.22%in F1-score compared to the direct input of low-light images.Additionally,image quality was further improved,as reflected in the enhancement of image information entropy,standard deviation,and average gradient by 0.62,13.24,and 21.61,respectively.The proposed model significantly outperformed similar unsupervised illumination models in average precision and F1-score enhancement in object detection tasks,without color distortion and overexposure in enhanced results.The proposed model can be widely applied to complex road-lighting scenarios,expanding the application of the existing defect detection,which is significant to application in engineering.