Blur is one of the common factors that deteriorate image quality and can be caused by various factors such as motion, defocus, or environmental conditions. The presence of partially or globally blurred images in a dataset can make object recognition challenging, thereby reducing the effectiveness of image classification models. To mitigate this issue, blurred images must either be removed from the dataset or processed using deblurring techniques. In this project, the impact of blurred images on the performance of deep learning-based image classification models investigated. Specifically, the goal was to analyze how different levels of image blur affect classification accuracy. To achieve this, a convolutional neural network (CNN) model was trained using the CIFAR-10 dataset, with varying proportions of blurred images: 0%, 25%, 50%, and 100%. The experiment results demonstrated that increasing the proportion of blurred images in the training dataset led to a decline in validation accuracy. The model achieved validation accuracies of 67.53%, 65.50%, 63.90%, and 55.74% when trained with datasets containing 0%, 25%, 50%, and 100% blurred images, respectively. These findings highlight the adverse effects of image blur on classification performance, emphasizing the importance of high-quality image data in deep learning applications.
Ahmet Fırat Yelkuvan (Mon,) studied this question.