Learning to Render Better Image Previews

Abstract

A rapidly increasing portion of Internet traffic is dominated by requests from mobile devices with limited-and metered-bandwidth constraints. To satisfy these requests, it has become standard practice for websites to transmit small and extremely compressed image previews as part of the initial page-load process. Recent work, based on an adaptive triangulation of the target image, has shown the ability to generate thumbnails of full images at extreme compression rates: 200 bytes or less with impressive gains (in terms of PSNR and SSIM) over both JPEG and WebP standards. However, qualitative assessments and preservation of semantic content can be less favorable. We present a novel method to significantly improve the reconstruction quality of the original image with no changes to the encoded information. Our neural-based decoding not only achieves higher PSNR and SSIM scores than the original methods, but also yields a substantial increase in semantic-level content preservation. In addition, by keeping the same encoding stream, our solution is completely inter-operable with the original decoder. The end result is suitable for a range of small-device deployments, as it involves only a single forward-pass through a small, scalable network.

Optimised Sparse Image Reconstruction Via Densification and Relocation

“…where the superscript denotes the iteration index. However, such a numerical descent estimation (8,9) is computationally expensive because it requires nine evaluations of the image reconstruction operator to estimate the descent direction for a single mask pixel.…”

“…The notation |P δ | denotes the number of mask pixels in the subset P δ . Compared to the single-pixel forward difference (8,9), the subset mask pixels P δ are perturbed together during the nine evaluations of the image reconstruction operator, while the remaining mask pixels P \ P δ are fixed. Therefore, the computational cost of the descent estimation reduces linearly as a function of the increasing block size |P δ |.…”

“…Reconstruction-based image/video compression [6,8,27,34,41,42,52,55] is a promising alternative to modern standard codecs, such as JPEG and HEVC that rely on transform-based compression. The main difference between reconstruction-based compression and transform-based compression is that the former works on the image domain without traditional transforms, such as Fig.…”

Accelerating Neural Network Inference on FPGA-Based Platforms—A Survey

“…Baluja et al [36] eliminates all multiplications and float-point operations by deploying a precomputed multiplication table. The authors receive |A| activations after quantizing nonlinear activation functions and confirm |W | weights in the neural network.…”

“…This approach does avoid the consumption of processing elements, but it presents new challenges for both storage and index by requiring scaling of the boundaries of the multiplication table in the face of different activation functions, and by calculating more multiplication tables if different quantization of weights is adopted. [36]. Calculations are transformed into search operations, which obviously reduces computation.…”