【作者】 刘康；

【导师】 潘刚； 吕耀志；

【作者基本信息】 天津大学 ， 工程（专业学位）， 2022， 硕士

【摘要】 目标检测作为经典的计算机视觉任务之一,是自动驾驶、智慧交通等领域的基石。但在有雾场景为代表的各种复杂天气条件下,其检测精度会因为训练数据与测试数据间差距的影响而显著下降。因此,提升目标检测算法在有雾条件下的稳定性是具有重要实际意义的研究方向。有雾场景中,物体成像因受到空气中微小粒子的影响而变得不清晰,从而影响目标检测的精度;另一方面,现有的去雾算法主要评价指标是PSNR、SSIM,而这些量化评价指标不能有效地评价去雾算法对于目标检测任务的提升。基于以上分析,本文为解决有雾场景下的目标检测做了以下三部分工作:(1)为了让去雾网络脱离PSNR等量化评价指标的约束真正帮助目标检测任务,本研究引入目标检测网络以感知损失的方式参与去雾网络的训练,以此帮助去雾网络在高级任务中获得准确度的提升。本文在实验中对比了使用感知损失和像素级损失的去雾结果,同时使用两种损失在PSNR和SSIM指标上取得最优,但仅使用感知损失的去雾模型在高级任务中取得更好的结果。(2)如果视目标检测网络连接去雾网络为一个整体,其已经可以解决有雾场景下的检测任务,但无雾场景中精度会下降。为了打破这种局限性,本文提出了Defog DA-Faster RCNN网络,其将领域适应引入到一体式网络中,使目标检测模块对于经过去雾模块的有雾、无雾两个域具有领域适应性。有雾图像经过去雾网络会得到更清晰的特征,无雾图像经过去雾网络带来的负面影响会被领域适应所弱化,领域适应与去雾网络相辅相成,充分发挥彼此的优势。(3)受到风格迁移等任务在Encoder-Decoder中间使用残差块来对表征进行变换的启发,本文提出了Defog FPN网络。其在FPN(Feature Pyramid Networks)的基础上嵌入了针对于特征进行去雾的Mini Defog模块,并且为了提升特征去雾的效果,本文还提出了一种一致性损失。经过实验证明基于Defog FPN的Faster RCNN网络检测结果超越了基于FPN的和连接去雾模块的目标检测网络。更多还原

【Abstract】 Object Detection is a common computer vision task,the detection accuracy of it will be significantly reduced due to the gap between training data and test data.At the same time,object detection is the cornerstone of automatic driving,intelligent transportation and other fields.In foggy scenes,the object imaging will become unclear due to the influence of small particles in the air,which will affect the accuracy of object detection.Improving the stability of object detection network under different weather is an important and practical research direction.In addition,the main goal of existing dehazing algorithms is high PSNR and SSIM,and such quantitative evaluation indicators cannot effectively evaluate the improvement of dehazing algorithms for object detection tasks.Based on the above analysis,this thesis makes efforts to solve the object detection task in foggy scenes from three directions:(1)In order to allow the dehazing network to break away from the constraints of quantitative evaluation indicators such as PSNR and truly help the object detection task,this study introduces the object detection network to participate in the training of the dehazing network in a perceptual-loss manner,so as to help the dehazing network improve the accuracy in high-level tasks.We compared the dehazing results using perceptual loss and pixel-level loss in our experiments,and using both losses at the same time achieves the best results in PSNR and SSIM metrics,but in the comparison of connected object detection tasks after image dehazing,using the defogging image with perceptual loss achieves better results.(2)If the object detection network is connected to the defogging network as a whole,it can solve the detection task in foggy scenes,but the accuracy will decline in fogless scenes.In order to break this limitation,this thesis proposes the Defog DAFaster RCNN network,which introduces domain adaptation into the integrated network,so that the object detection module has domain adaptability for both foggy and fogfree domains after the dehazing module.The negative impact of the haze-free image through the dehazing network will be digested by domain adaptation.For the hazy image,it will get clearer features through the dehazing network.The domain adaptation and the dehazing network complement each other and perform better than before.In experiments we prove that the network can achieve excellent performance in foggy and fog-free scenes.(3)Inspired by the using residual blocks in the middle of Encoder-Decoder to transform representations in tasks such as style transfer,we propose Defog FPN.On the basis of Feature Pyramid Networks,it embeds the Mini Defog network for feature dehazing,and in order to improve the effect of feature dehazing,we propose a consistency loss.Experiments demonstrate that the detection results of Faster RCNN network based on Defog FPN outperform than others.更多还原