基于深层特征学习的可压缩感知及缝雕刻的图像重定向

Compressive perception and seam carving image reretargeting based on deep feature learning

多媒体技术的飞速发展推动了图像处理与显示设备的应用与发展,为了使图像在不同的设备上进行最佳显示,需要对图像的尺寸进行调整。因此,本文提出一种基于深层特征学习的可压缩感知及接缝雕刻的图像重定向方法。首先从预先训练的VGG-19网络中提取输入图像的深度特征图,从最深层开始计算特征图像的可压缩率,根据计算的可压缩率运用接缝雕刻的方法在特征域(Feature fields Seam Carving,FSC)调整特征图的大小,然后依次向较浅的层传播,得到所有特征层的重定向图像后,将输入图像对应于第一层特征图的去缝的位置处的像素去掉,得到原始图像的重定向图像。若没有达到目标图像的大小,最后再进行均匀缩放(scaling,SCL)。在RetargetMe数据集上分别进行主观与客观评估,结果表明,与其他方法相比,本文的重定向方法总体上实现了更好的性能。

The rapid development of multimedia technology has promoted the application and development of image processing and display devices.In order to make images optimally displayed on different devices,it is necessary to adjust the size of the image.Therefore,this paper proposes an image retargeting method based on compressible sensing and seam carving based on deep feature learning.First,extract the feature map of the input image from the pre-trained VGG-19 network,calculate the compressibility of the feature image from the deepest layer,and apply the seam carving method on feature fields(FSC)according to the calculated compressibility to adjust the size of the feature map,and then propagate to the shallow layer in turn to obtain the retargeted image of all the feature layers,and then the pixels at the positions where the input image corresponds to the unstitched portion of the first layer feature image are removed,and a retargeted image of the original image is obtained.If the size of the target image is not reached,the final scaling is performed.Subjective and objective evaluations were performed on the RetargetMe dataset respectively.The results show that,compared with other methods the retargeting method of this paper achieves better performance overall.