MNIST-C
收藏资源简介:
MNIST-C是由谷歌公司创建的一个用于计算机视觉领域分布外鲁棒性测试的数据集,包含15种不同的图像损坏方式。该数据集源自MNIST测试集,通过应用多种损坏如噪声、模糊、几何变换等,旨在评估模型在面对非理想图像输入时的性能。创建过程中,从31种潜在的损坏中精选出15种,确保每种损坏都能有效降低模型准确率同时保持图像的语义内容。MNIST-C的应用领域主要集中在提高模型对图像损坏的鲁棒性,解决模型在实际应用中可能遇到的图像质量问题。
MNIST-C is a dataset created by Google for out-of-distribution robustness testing in the field of computer vision, containing 15 distinct types of image corruptions. Derived from the MNIST test set, this dataset applies various corruptions such as noise, blurring, geometric transformations and others, aiming to evaluate model performance when confronted with non-ideal image inputs. During its creation, 15 types were selected from 31 potential corruptions, ensuring that each corruption can effectively reduce model accuracy while preserving the semantic content of the images. The main application areas of MNIST-C focus on improving the robustness of models against image corruptions, addressing image quality issues that models may encounter in real-world practical applications.




