Image Quality Assessment via Inter-class and Intra-class Differences for Efficient Classification

With the development of data-centric artificial intelligence, more and more people pay attention to the importance of image information quality. Based on the core idea that images in datasets have different intra-class information richness and inter-class information overlaps, we propose a two-stage image quality assessment method. The images in the area with both a lower intra-class richness and a higher degree of inter-class overlap can provide more image information for the neural network, thus further improve the model performance. Experiments on two public image classification datasets for image classification (CIFAR10 and mini-ImageNet) show that the proposed image information quality assessment method can effectively distinguish high information quality images. Under the same budget, selecting images with higher image information quality can achieve better performances than lower image information quality (Testing accuracy: 1.69% higher on CIFAR10, 2.11% higher on mini-ImageNet).