用于视频图像帧间运动补偿的深度卷积神经网络

为探索深度学习理论在视频图像帧间运动补偿问题中的应用，提出一种用于视频图像帧间运动补偿的深度卷积神经网络。该网络由卷积模块和反卷积模块构成，可以处理不同分辨率输入图像并具备保持较完整图像细节的能力。利用具有时序一致性的视频图像序列构造训练样本，采用随机梯度下降法对设计的深度卷积神经网络进行训练。视觉效果和数值评估实验表明，训练得到的网络较传统方法能更有效地进行视频图像帧间运动补偿。
In order to explore the application of deep learning theory in the problem of motion compensated frame interpolation, a DCNN (deep convolutional neural network) built with convolutional blocks and deconvolutional blocks was proposed. The proposed DCNN is capable of processing input images with different resolutions and preserving fine grained image details. The temporal coherent image sequences were used to construct the training sample and the stochastic gradient descent method was adopted to train the designed DCNN. Qualitative and quantitative experiments show that the trained DCNN obtains better interpolated images than the traditional approach in two testing images sequences.