[Paper Review] A Deeper Look into Hybrid Images

Jimut Bahan Pal|arXiv (Cornell University)|Jan 30, 2020

Advanced Image Fusion Techniques2 references4 citations

TL;DR

This paper investigates the creation and optimization of hybrid images—static images that appear as one image when viewed up close and another when viewed from afar—by combining low-pass and high-pass filtered images using Gaussian and Laplacian of Gaussian filters. The key contribution is empirical analysis of parameter trade-offs, particularly sigma values and blending weights, showing that optimal results require manual tuning and are constrained by computational cost, limiting real-time applicability.

ABSTRACT

$Hybrid$ $images$ was first introduced by Olivia et al., that produced static images with two interpretations such that the images changes as a function of viewing distance. Hybrid images are built by studying human processing of multiscale images and are motivated by masking studies in visual perception. The first introduction of hybrid images showed that two images can be blend together with a high pass filter and a low pass filter in such a way that when the blended image is viewed from a distance, the high pass filter fades away and the low pass filter becomes prominent. Our main aim here is to study and review the original paper by changing and tweaking certain parameters to see how they affect the quality of the blended image produced. We have used exhaustively different set of images and filters to see how they function and whether this can be used in a real time system or not.

Motivation & Objective

• To analyze the impact of varying sigma values and blending weights on the perceptual quality of hybrid images.
• To investigate the computational cost of generating hybrid images using low-pass and high-pass filters.
• To evaluate the feasibility of hybrid image generation in real-time systems.
• To explore parameter tuning strategies for achieving clear, distance-dependent visual transitions between two images.
• To assess the role of image contrast and color in the visibility of hybrid image components.

Proposed method

• The study applies a Gaussian low-pass filter and a Laplacian of Gaussian (LoG) high-pass filter to two source images with varying sigma values (e.g., σ = 10, 20).
• The filtered images are combined using weighted addition, with the low-pass image weighted at 0.85 and the high-pass at 0.15 for optimal results in one case study.
• An automated script measures computation time and image size across multiple kernel sizes (2 to 30) and image dimensions.
• The system was tested on a machine with specific hardware (Figure 4.1), and performance was logged for time vs. image size.
• The researchers used a trial-and-error approach to determine optimal blending weights, especially for images with contrasting colors.
• The method involves convolution-based filtering using Weierstrass transformation and 2D Gaussian functions for spatial frequency separation.

Experimental results

Research questions

• RQ1 How do different sigma values affect the clarity and perceptual transition of hybrid images?
• RQ2 What is the optimal blending weight ratio between low-pass and high-pass filtered images for clear visual perception at different viewing distances?
• RQ3 How does image size and filter kernel size influence computation time in hybrid image generation?
• RQ4 To what extent do color contrast and image content affect the visibility of the two interpretations in a hybrid image?
• RQ5 Can hybrid image generation be made efficient enough for real-time applications with current computational methods?

Key findings

• A sigma value of 10 produced a balanced hybrid image for Swami Vivekananda (low-pass) and Ramakrishna (high-pass), with optimal blending at 0.85× low-pass and 0.15× high-pass weights.
• Images of size 1500×1500 took approximately 1000 seconds to process with σ = 30, indicating high computational cost.
• Low-pass filtering took longer than high-pass filtering, possibly due to caching or NumPy optimization effects.
• The study found no automatic method for determining optimal blending weights; manual tuning via trial and error was required.
• There is a trade-off between sigma value and blending weight: higher sigma increases blurriness, requiring more weight on the low-pass image to remain visible from a distance.
• The method is not feasible for real-time systems due to excessive computation time, especially for large images and high sigma values.