Multi-operator image retargeting in compressed domain by preserving aspect ratio of important contents

Content-aware image retargeting have received extensively research attentions. However, most of exiting retargeting approaches perform resizing on raw image data in the pixel domain. Since images in the actual world are mostly stored and transmitted in the compressed domain, e.g. discrete cosine transformation (DCT) domain, the complete decompression and recompression are almost inevitable by using the pixel domain based retargeting methods, causing extra overheads with high computation complexity. To address this issue, we propose a novel multi-operator image retargeting method in the DCT domain, in which three techniques including indirect seam carving, similarity transformation, and direct seam carving based on gradient vector flow (GVF), are utilized to perform resizing. To eliminate the zigzag effects in the retargeted images, we also present a novel similarity transformation algorithm in the DCT domain by which the DCT coefficients instead of a whole block are rescaled during resizing. In addition, we develop two decoding schemes to solve the issue that the traditional inverse DCT cannot be directly applied to the decoding the retargeted images. Extensive results demonstrate that the presented multi-operator image retargeting method in the DCT domain can preserve the aspect ratio of visual important contents well and obtain the resized images of better quality than the existing methods.