Image Transformation for Perceptual Size Restoration

Abstract

The thing of interest(ToI) in the photo may appear smaller than its actual size due to the discrepancy between the image processing principles of the camera and human perception. This is because the camera produces photos using perspective projection, while human perception is affected by the observation location and psychological factors. When conventional image resizing methods are applied to enlarge the ToI of the input image, they cause problems that impair the perceptual impression, such as loss of depth perception, distortion of composition, and deformation of salient objects. Therefore, in this study, we propose an image transformation method based on seam carving to enlarge the ToI size. The proposed method defines a new energy function to maintain the position of the ToI in the enlarged image. Moreover, we employ the state-of-the-art deep learning models to detect the ToI and the salient regions. Then we modify the seam carving algorithm to achieve the enlargement of the ToI. We conducted objective and subjective evaluation experiments to validate the effectiveness of the proposed method. The objective evaluation experiment demonstrated that the defined compositional energy was effective. The subjective evaluation experiment indicated that the proposed method was effective in reproducing the perceived size and preserving the depth perception of the ToI.