Performance Evaluation of Deep Pre-Trained Models Under Progressive Blur

In this paper, we evaluate the feature extraction capability and classification performance of five contemporary deep pre-trained models on progressively blurred images of handwritten digits from the MNIST handwritten digits dataset. The images are progressively blurred, first using a Gaussian blur with Sigma=5, and then using a Gaussian blur with Sigma=8. Sigma is also known as standard deviation; higher the Sigma, larger the amount of blur. The deep pretrained models under consideration are VGG-16, DenseNet121, Xception, ShuffleNet, and SqueezeNet, that are pretrained on the ImageNet dataset, and subsequently shallowtuned on the blurred images. Among the deep learning models, DenseNet-121 achieved the highest accuracy of 98.77% for Sigma=5 and an accuracy of 98.62% for an increased blur of Sigma=8. All other model accuracies, with the exception of ShuffleNet, significantly declined when the amount of blur increased (Sigma is increased from 5 to 8). Comparisons with the machine learning models - Support Vector Machine (SVM), Convolutional Neural Network (CNN), and logistic regression, prove that the performance of logistic regression is far superior to other machine learning models, though the accuracies were lower than most of the deep learning models. We conclude that DenseNet-121, followed by ShuffleNet, are some of the best contemporary models that can classify accurately under progressive blur.