恰可察觉QP(JNQP)数据

通过测试人员对测试视频的主观质量评价，获得三个质量层级的JNQP数据。量化人眼对编码视频质量层级的非线性感知行为，从而做到更精确的恰可察觉编码失真（JNCD）计算。优化视频编码环节，在不降低视频质量的同时减少编码编码数据量，提升编码效率。1.数据采集：主观测试在杭州电子科技大学丽水研究院进行，每次测试都是在背景亮度和色度相同的非分散注意力实验室环境下，观看设备为 HP P27q G4 2K显示器。2.数据采集算法：在主观测试中，每个受试者需要相继比较连续播放的两段视频的质量，并确定这两个视频序列的感知质量是否明显不同，受试者应选择“是”或者“否”。首先初始搜索范围为（，），其中=0，=51。第一次搜索将源视频设置为参考视频，与 QP 为 (+)/2的压缩视频进行比较，如果受试者给出“否”的答案，即两个视频在感知上没有区别，这表明 JND 点在 QP 值为 (+)/2的前面，此时新的搜索范围（，）。反之，如果受试者给出“是”的答案，则搜索范围变为（，）。到第二次搜索时，将参考视频与QP为搜索范围中值的压缩视频比较，直到最终搜索范围小于等于 1 时，流程结束。测试后续质量层级的JNQP点时，参考视频与初始范围中的选取上一质量层级的QP点，=51。3.视频数据处理：对采集到数据进行错误样本剔除、离群样本剔除、不可信受试者样本剔除，减少误差。JNQP data with three quality levels were obtained via subjective quality evaluation of test videos by testers. This work quantifies the nonlinear perceptual behavior of the human eye regarding encoded video quality tiers, enabling more accurate calculation of Just Noticeable Coding Distortion (JNCD). It optimizes the video coding pipeline, reducing coded data volume without sacrificing video quality and improving coding efficiency. 1. Data Collection: The subjective tests were conducted at the Lishui Research Institute of Hangzhou Dianzi University. Each test was carried out in a distraction-free laboratory environment with consistent background luminance and chrominance, using an HP P27q G4 2K monitor as the viewing device. 2. Data Collection Algorithm: During the subjective test, each participant was required to sequentially compare the quality of two consecutively played video clips and determine whether the perceived quality of the two video sequences differed significantly, by selecting either "Yes" or "No". Initially, the search range was set as (QLow, QHigh), where QLow=0 and QHigh=51. For the first search, the source video was used as the reference video, and it was compared with the compressed video with QP = (QLow + QHigh)/2. If the participant answered "No", meaning the two videos were perceptually indistinguishable, this indicated that the JND point was below the QP value of (QLow + QHigh)/2, and the new search range was updated to (QLow, (QLow + QHigh)/2). Conversely, if the participant answered "Yes", the search range was adjusted to ((QLow + QHigh)/2, QHigh). For the second search, the reference video was compared with the compressed video with QP equal to the median of the current search range. This process was repeated until the final search range was less than or equal to 1, at which point the procedure terminated. When testing JNQP points for subsequent quality tiers, the reference video adopted the QP value of the previous quality level selected from the initial range, with QHigh=51. 3. Video Data Processing: Error samples, outlier samples, and samples from unreliable participants were removed from the collected data to minimize measurement errors.

搜集汇总

数据集介绍

特点

恰可察觉QP(JNQP)数据是一个用于量化人眼对编码视频质量层级感知行为的数据集，包含577条记录，主要用于优化视频编码环节，提升编码效率。