| 结合多尺度特征融合和多输入多输出编-解码器的去模糊算法 |
| |
| 引用本文： | 赵倩,周冬明,杨浩,王长城,李淼.结合多尺度特征融合和多输入多输出编-解码器的去模糊算法[J].红外与激光工程,2022,51(10):20220018-1-20220018-13. |
| |
| 作者姓名： | 赵倩 周冬明 杨浩 王长城 李淼 |
| |
| 作者单位： | 云南大学 信息学院，云南 昆明 650504 |
| |
| 基金项目： | 国家自然科学基金(62066047, 61966037) |
| |
| 摘 要： | 针对相机抖动、拍摄物体快速运动以及低快门速度等因素造成的图像非均匀模糊，提出一种结合多尺度特征融合和多输入多输出编-解码器的去模糊算法。首先使用多尺度特征提取模块来提取较小尺度模糊图像的初始特征，该模块使用扩张卷积来以较少的参数量获得更大的感受野。其次，通过特征注意力模块来自适应地学习不同尺度特征中的有效信息，该模块利用小尺度图像的特征来生成注意图，能够有效地减少冗余特征。最后，使用多尺度特征渐进融合模块逐步融合不同尺度的特征，使得不同尺度特征信息能够进行互补。相比以往的使用多个子网堆叠的多尺度方法，文中使用单个网络就能提取多尺度特征，从而降低了训练难度。为了评估网络的去模糊效果和泛化性能，提出的算法在基准数据集GoPro、HIDE和真实数据集RealBlur上均进行了测试。在GoPro和HIDE数据集上的峰值信噪比值分别为31.73 dB和29.39 dB，结构相似度值分别为0.951和0.923，其结果均高于目前先进的去模糊算法，并且在真实数据集RealBlur上也取得了最佳效果。实验结果表明，提出的去模糊算法相比现有算法去模糊更为彻底，能有效地复原图像的边缘轮廓和纹理细节信息，并且能够提升后续高级计算机视觉任务的鲁棒性。
|
| 关 键 词： | 图像去模糊 图像恢复 深度学习 多输入多输出 多尺度网络 |
| 收稿时间： | 2022-01-06 |
Image deblurring via multi-scale feature fusion and multi-input multi-output encoder-decoder |
| |
| Affiliation: | School of Information Science & Engineering, Yunnan University, Kunming 650504, China |
| |
| Abstract: | A deblurring method combining multi-scale feature fusion and a multi-input multi-output encoder-decoder is proposed for non-uniform blurred images caused by camera shake, fast motion of the captured object, and low shutter speed. Firstly, the initial features of smaller-scale blurred images are extracted using a multi-scale feature extraction module, which uses dilated convolution to obtain a larger receptive field with a smaller number of parameters. Second, the feature attention module is used to adaptively learn useful information from different scale features, which can effectively reduce redundant features by using features of small-scale images to generate attention maps. Finally, the multi-scale feature progressive fusion module is applied to gradually fuse features at different scales, making the information of different scale features to complement each other. Compared with recent multi-scale methods that use multiple subnets stacked on top of each other, we use a single network to extract multi-scale features, thus reducing the training difficulty. To evaluate the deblurring effect and generalization performance of the network, the proposed method is tested on both the benchmark datasets GoPro, HIDE, and the real dataset RealBlur. The peak signal-to-noise ratio values of 31.73 dB and 29.39 dB and the structural similarity values of 0.951 and 0.923 on the GoPro and HIDE datasets, respectively. The deblurring performance is higher than that of recent state-of-the-art deblurring methods, and it also has better performance on the RealBlur dataset containing real scenarios. The experimental results demonstrate that the proposed method is more effective than recent deblurring methods, can effectively restore the edge contour and texture detail information of images. In addition, our method can improve the robustness of subsequent high-level computer vision tasks. |
| |
| Keywords: | |
|
| 点击此处可从《红外与激光工程》浏览原始摘要信息 |
|
点击此处可从《红外与激光工程》下载全文 |
|
