高级检索

聚合双重上下文特征的车辆玻璃图像反射去除

Aggregating Dual Contextual Features for Vehicle Glass Image Reflection Removal

  • 摘要: 在智慧交通领域中, 车辆玻璃的反射现象可能导致监控抓拍的车内图像出现模糊、变形或遮挡等情况, 进而形成带反射的混合图像, 直接影响监控系统对车内信息的准确判断和识别. 针对现有的图像反射去除方法大都没有考虑特征聚合, 极易造成颜色失真、伪影明显等问题, 提出一种聚合双重上下文特征的车辆玻璃图像反射去除方法. 首先创建多尺度特征融合的特征金字塔网络, 提取混合图像的空间上下文特征, 获取所需的空间信息来改善伪影明显现象; 然后创建引入特征注意力机制的编解码器, 提取混合图像的通道上下文特征, 获取所需的通道信息来改善颜色失真现象; 最后构建特征级聚合网络, 依托像素注意力机制引导特征聚合, 实现车辆玻璃图像反射去除的同时有效地提升背景图像的可识别性. 在Real20和SIR²测试数据集上的实验结果表明, 所提方法的峰值信噪比、结构相似度和最小均方误差均优于现有图像反射去除方法, 伪影明显和颜色失真等问题得到明显改善, 证明了该方法在图像反射去除中的有效性.

     

    Abstract: In the field of intelligent transportation, the reflection phenomenon of vehicle glass may lead to blurring, distortion, or blocking of vehicle interior images captured by the surveillance system, thus forming a mixed image with reflections, which directly affects the surveillance system's accurate judgment and recognition of vehicle interior information. In view of the fact that most of the existing image reflection removal methods do not consider feature aggregation, which can easily cause color distortion and obvious artifacts, we propose a method to remove reflections from vehicle glass images by aggregating dual contextual features. Firstly, a multi-scale feature fusion of feature pyramid network is created to extract the spatial context features of the hybrid image, and obtain the required spatial information to improve the phenomenon of obvious artifacts; then an encoder-decoder structure is created that introduce the feature attention mechanism, and the channel context features of the hybrid image are extracted, and the required channel information is obtained to improve the color distortion phenomenon; finally, a feature-level aggregation network is constructed, which relies on the pixel attention mechanism to guide the feature aggregation, and realizes the removal of reflections from the vehicle glass image while effectively improving the recognizability of the background image. The experimental results on Real20 and SIR² test datasets show that the peak signal-to-noise ratio, structural similarity and minimum mean square error of the proposed method outperform those of existing reflection removal methods, and the problems of artifacts and color distortions have been significantly improved, which proves the effectiveness of the proposed method in the removal of image reflections.

     

/

返回文章
返回