Abstract:

Aiming at the problem of treating uncertainty modeling as a step for optimizing prediction results while ignoring the value of uncertainty itself in current object detection algorithms， an object detection result evaluation scheme based on Guide to the Expression of Uncertainty in Measurement （GUM） was proposed. Firstly， sources of uncertainty in object detection were decomposed into three mutually independent aspects： data， model and platform. Then， uncertainty influence factors were extracted from these three aspects to construct measurement function. Secondly， type A and type B evaluation methods in GUM were used to measure uncertainty influence components. Finally， uncertainty synthesis rules were used on the basis of the measurement function， and standard uncertainty was synthesized from uncertainty components. Experiments were conducted by using an object detection algorithm. The results show that compared to Peak Signal-to-Noise Ratio （PSNR） and Structural SIMilarity （SSIM）， the data uncertainty increased 5.30 and 19.08 percentage points respectively in capturing noisy data； model uncertainty has a tiny influence on the prediction results， which can be ignored within the range of 10^-6； platform uncertainty can represent prediction result differences caused by software and hardware platforms in numerical form.

Key words: object detection algorithm, uncertainty measurement, uncertainty decomposition and synthesis, image data, software and hardware platforms

摘要：

针对当前目标检测算法的不确定性研究将不确定性建模作为优化预测结果的一个步骤，而忽视了不确定性本身的性质的问题，提出一种基于测量不确定度表示指南（GUM）的目标检测结果评定方案。首先，将目标检测的不确定度来源分解为数据、模型、平台3个互相独立的方面，并从这3个方面提取不确定度影响因素，从而构建不确定度测量函数；其次，使用GUM中的A类评定方法和B类评定方法对不确定度影响分量进行度量；最后，基于测量函数使用不确定度合成法则，并由不确定度分量合成标准不确定度。在目标检测算法上展开实验，结果表明，与峰值信噪比（PSNR）和结构相似性（SSIM）相比，数据不确定度在捕捉噪声数据方面分别提高了5.30和19.08个百分点；模型不确定度对预测结果的影响很小，在10^-6范围内可以忽略；平台的不确定度可以用数值化形式表示由软硬件平台带来的预测结果差异。

关键词: 目标检测算法, 不确定度测量, 不确定度分解与合成, 图像数据, 软硬件平台