How JD Cloud’s Mobile Super‑Resolution SDK Boosts Video Quality and Cuts Bandwidth by 30%

Overview

Video super‑resolution (super‑resolution) uses deep learning to analyze video frames and scenes, applying denoising, de‑blurring, sharpening, and de‑jittering to improve visual quality while reducing production and transmission costs.

Leveraging JD Cloud’s expertise in video codec technology, algorithm optimization, and assembly‑level enhancements, the newly released mobile super‑resolution SDK for Android and iOS has been deployed in the JD Mall app, increasing average playback duration by 80% and lowering bandwidth cost by 30%, thereby boosting user experience and GMV conversion.

Technical Implementation

Existing image and video super‑resolution methods fall into single‑image super‑resolution (SISR) and video super‑resolution (VSR). For real‑time live streaming, JD Cloud adopts the linear network ESPCN algorithm, optimized for ROI characteristics, and adds engineering improvements to handle artifacts such as ringing, block effects, text blur, and video jitter.

ESPCN moves up‑sampling after convolution, extracting features in the low‑resolution (LR) space and using sub‑pixel convolution for up‑sampling. The video is divided into ROI and non‑ROI slices; ROI regions are processed with ESPCN, while non‑ROI regions use traditional bicubic up‑sampling. The slices are then stitched to reconstruct the full YUV/RGB frame, enabling real‑time 1080p‑to‑4K super‑resolution on modest hardware.

Overall Architecture

The human visual system is highly sensitive to luminance, so super‑resolution focuses on reconstructing the luminance channel, while chroma channels are up‑sampled using conventional interpolation.

Processing steps: decode low‑resolution video, separate luminance and chroma signals, enhance each, apply ESPCN‑based super‑resolution to the luminance, bicubic up‑sampling to chroma, merge channels, convert YUV to RGB, and output the high‑resolution video.

Technical Application

The SDK has been integrated into the JD Mall app, showing significant visual improvements.

Evaluation Results

Objective metrics comparing the super‑resolution method with traditional approaches and the original source are as follows:

Resolution 540×960 – PSNR: 33.43 / 30.21, VMAF: 92.33 / 78.16, SSIM: 0.918 / 0.832.

Resolution 424×430 – PSNR: 32.55 / 30.43, VMAF: 93.46 / 79.47, SSIM: 0.924 / 0.896.

Power‑consumption and performance comparisons also demonstrate the SDK’s efficiency.

Product Overview

As a leading video cloud provider in China, JD Cloud Video offers live streaming, video‑on‑demand, real‑time communication, and full‑stack SDKs, covering the entire video pipeline from capture to storage, management, and distribution. Its breakthroughs in encoding, algorithm optimization, and audio‑video processing give it a competitive edge in ultra‑high‑definition transcoding, low‑latency streaming, and end‑to‑end video solutions.