| 摘要: |
| 现有的低光照图像目标识别方法通常将图像恢复与目标识别任务分开处理，且在图像恢复质量和计算时间上无法满足目标识别任务的要求.针对这些问题，首先，提出一种高效的图像恢复卷积神经网络架构，通过融合不同尺度的特征图来聚合多级上下文特征信息，减少卷积层的信息冗余，提高图像恢复任务的实时性.并且设计了一个局部-全局注意力模块，通过校准各特征图的局部信息和特征通道之间的关系，提高恢复网络对噪声和图像内容的区分能力.其次，提出了一种图像恢复和目标识别任务协同处理的方法.利用目标识别的高级语义信息指导图像恢复网络学习，从而突出目标的结构和纹理等特征，使恢复结果更加适应目标识别任务.实验结果表明，该方法在图像恢复质量、计算时间和目标识别率上明显优于已有方法. |
| 关键词: 深度学习 卷积神经网络 目标识别 低光照图像恢复 注意力机制 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金（No.61370222）；黑龙江省自然科学基金（ZD2019F003） |
|
| Methods of Image Restoration and Object Detection in Low-Light Environment |
|
REN Dong-Dong1,2, LI Jin-Bao1,2
|
|
1.School of Computer Science and Technology, Heilongjiang University, Harbin 150080, China;2.Heilongjiang Provincial Key Laboratory for Database and Parallel Compluting(Heilongjiang University), Harbin 150080, China
|
| Abstract: |
| The existing object detection methods for low-light image usually separate image restoration from object detection tasks. In addition, the quality and computing time of image restoration cannot meet the requirements of object detection task. To solve these problems, firstly, this study proposes an efficient image restoration convolutional neural network architecture, which aggregates feature information of multi-level contexts by combining feature maps of different scales, reduces information redundancy of convolutional layers, and improves the real-time performance of image restoration. In addition, a local-global attention block is designed to improve the ability of the recovery network to distinguish between noise and image content by calibrating the local information of each feature map and the relationship between feature channels. Secondly, this study designs a solution for collaborative processing of image restoration and target recognition tasks. The high-level semantic information of target recognition is used to guide the image recovery network learning, so as to highlight the feature information such as the structure and texture of the target, and make the recovery result more suitable for the target recognition task. Experimental results show that this method is superior to the existing methods in image restoration quality, computing time and object detection rate. |
| Key words: deep learning convolutional neural network object detection low-light image reatoration attention mechanism |
