Multi-focus image fusion using Singular Value Decomposition in DCT domain

Multi-focus image fusion is a process that fuses several images from a scene with different focal lengths into a whole image in which all areas are focused on. Image fusion methods in the Discrete Cosine Transform (DCT) domain are efficient due to their low time and energy consumption, and low complexity. This is especially true when fusing images are compressed in JPEG format in Visual Sensor Networks (VSN). In this paper, a low complexity multi-focus image fusion in DCT domain is presented which increases the output image quality. Our proposed method makes it suitable for real-time applications because of its implementation in DCT domain. On the other hand, it is stable in noisy conditions. The proposed method uses the singular values of Singular Value Decomposition (SVD) of 8×8 input blocks in DCT domain. The geometric mean of the 5 largest singular values (out of 8 singular values) is computed as a criterion of focused block detection. The blocks which have the highest geometric mean value among other corresponding blocks is selected as the focused block. These blocks are then used for constructing the output image. This method can be utilized both in DCT domain and in spatial domain. Various experiments and comparisons between the proposed method and the previous methods in noisy and noiseless conditions have been presented, which confirm the increase in image quality and stability in noisy images.

I. Introduction

The process of image fusion consists of combining multiple images to obtain a single image that is more precise than the original images. An aim of the image fusion is to decrease the information quantity, while simultaneously making it more interpretable for both human and machine perception. In multi-focus image fusion case, the single output fused image is constructed from the multiple input images which each of them is focused on a part of scene and the other parts are blurred [1], [2]. Technically speaking, due to the limited depth of focus in optical lenses of CMOS/CCD cameras, recording an image which has distinct components is challenging [3]. In order to create a vivid image, every part of image must remain in the focal length. However, this is usually not practical and only a fraction of the image, which is remaining in the focal length could be explicit, leaving the rest obscure [4], [5]. In Visual Sensor Networks (VSN) there is an opportunity to capture more images from a scene by multi cameras with different depths of focus [5]–[8].