A phase detection autofocus (PDAF) algorithm iteratively estimates the phase shift between the left and right phase images captured in an autofocus process and uses it to determine the lens movement until the estimated in-focus lens position is reached. Such phase images have been assumed to be equivalent to a two-view light field. If the assumption is true, then the phase shift between the two phase images can be obtained by stereo matching or similar techniques. In this paper, we argue that it is a wrong assumption and provide insights into the distinctions between phase images and two-view light field from the autofocus perspective. We also support our argument by conducting an experiment to show that both stereo matching and optical flow result in inferior PDAF performance than the phase correlation technique and the AF-Net technique that specifically target phase images.
In recent years, smartphones have become the primary device for day-to-day photography. Therefore, it is critical for mobile imaging to capture sharp images automatically without human intervention. In this paper, we formulate autofocus as a decisionmaking process, in which the travel distance of a lens is determined from the phase data obtained from the phase sensors of a smartphone, and the decision-making policy is based on reinforcement learning, a popular technique in the field of deep learning. We propose to use a noise-tolerant reward function to combat the noise of the phase data. In addition, instead of using only the current phase data, each lens movement is determined using the phase data acquired along the journey of an autofocus process. As a result, the proposed machine-learning approach is able to expedite the autofocus process as well. Experimental results show that the method indeed improves the autofocus speed.