ABSTRACT
Image blurring causes loss of information and restoring the degraded image requires knowledge about the original image and the blur kernel, which is usually unknown. However, if information about the blur kernel is available, the image can be deblurred. Degradation due to defocus blur is considered in this paper. To estimate the blur kernel and facilitate in the process of restoration, determination of amount of defocus radius is essential. Here, a convolutional neural network architecture is proposed to determine the defocus blur radius. The Fourier spectrum of defocus blurred images is fed as an input to the network for extraction of the features and classification of the defocus radius. The network is trained, validated and tested using images from standard datasets like USC-SIPI, Berkeley Segmentation dataset and Pascal VOC 2007. The experimental results show competitive results as compared to existing defocus radius estimation methods.
