Mining and reasoning of data uncertainty-induced imprecision in deep image classification

Abstract

Existing deep image classification techniques strive to suppress data uncertainty for various reasons such as blur, occlusion, noise, and label error and advance to higher accuracy. However, they ignore data uncertainty-induced imprecision and thus do not work as intended. In this paper, we propose a deep open-source framework to mine and reason the imprecision of such training and test sets, which can present better performance and reduce the risk of misclassification. First, we design a label reassignment mechanism. It allows the network to reassign training labels and allow imperfect training samples with multiple labels. As a result, they are removed from the original classes and considered new imprecise samples to represent partial ignorance. Second, we propose a new imbalanced data enhancement architecture to learn a generalized representation of each (precise and imprecise) class. It helps the network fuse the auxiliary information from both precise and imprecise classes, which is beneficial to extract more distinctive class features from single-labeled samples and characterize uncertainty-induced imprecision in the test set by imprecise test samples. Afterward, methodological analyses and empirical evaluations are conducted. The proposed framework is demonstrated to present better performance on different typical networks (Resnet50, MobileNetV2, DenseNet121, EfficientNetB0, ShuffleNetV2, SENet, SqueezeNet, Xception) based on five publicly available datasets (Imagewoof-5, Flowers, Monkey, Butterfly, and Cifar-10). In addition, several targeted deep techniques for uncertain images or imprecise results are also employed as comparisons to prove the superiority of the proposed framework.