Game Development Reference
In-Depth Information
demand of compressing the local features for low bitrate visual search. To reduce
the bandwidth, efforts have been devoted to develop compact global or local feature
descriptors, such as Chen et al. (
2009
), Ji et al. (
2012
) andChandrasekhar et al. (
2009
).
Moreover, recent endeavors of Compact Descriptor for Visual Search (CDVS)MPEG
standardization also focus on the issue of interoperability (Duan et al.
2013
), that is,
allowing effective and efficient comparison of compact descriptors without decoding.
The main issue in the context of compression with feature descriptors is to recon-
struct the input image with image local descriptors like SIFT. In Weinzaepfel et al.
(
2011
), the authors proposed to stitch the patches from the image database. In Eitz
et al. (
2011
), an interactive system to synthesize images from sparse user sketches is
proposed. These approaches have presented novel ideas on visual signal representa-
tion and casted new light on high efficiency image compression.
The representative framework of cloud image compression was proposed in Yue
et al. (
2013
), in which the input image is described with the sub-sampled images
and SIFT features. As illustrated in Fig.
12.7
, the input image is first downsampled
and compressed with conventional image compression schemes, and then transmit-
ted as the base layer. To reconstruct high-quality image at the receiver, local features
descriptor SIFT is extracted from the original image. The correlation of the local fea-
tures between the original and sub-sampled images is further exploited and only the
residuals of the feature vectors are transmitted. The residuals of SIFT features are then
subject to transformation, quantization and entropy coding for efficient compression.
At the decoder side, the decoded image is upsampled to the original resolution. Then
the prediction SIFT descriptors are extracted from the subsampled image, and the
final SIFT descriptors are reconstructed by adding the SIFT residual and prediction
values. To reconstruct the original image, reconstructed SIFT descriptors are used
to retrieve high-quality patches in the cloud, and for each patch, it is stitched to the
upsampled decompressed image for high-quality reconstruction.
Experimental results show that compared to the state-of-the-art image compres-
sion algorithm such as HEVC intracoding, the cloud-based image coding scheme can
achieve significantly better reconstruction visual quality with much less bitrate. The
proposed idea brings image coding from internal redundancy reduction to external
reduction step. However, how to ensure the pixel-wise fidelity remains unresolved
in this feature-based compression framework.
Fig. 12.7
Encoder framework in cloud-based image compression
