Introduction
In the mid-1990s, psychophysical experiments on high-level vision, such as object recognition, were largely restricted to abstract stimuli like line drawings or simple geometric shapes. While researchers aimed to use realistic gray-scale images, they were hindered by technical difficulties in manipulating these images without introducing unintended visual artifacts. Furthermore, there was theoretical uncertainty regarding which image features were most relevant to the brain's processing of complex scenes. This paper introduces a software methodology to overcome these hurdles by creating "complementary" gray-scale images—pairs that share no common visual information but together form a complete picture—using Gabor wavelet decomposition.
Methodology: Gabor Wavelet Decomposition
The core of this approach is the use of Gabor functions, which are highly valued in vision science for their ability to model the receptive fields of neurons in the primary visual cortex (V1). These functions act as localized, oriented band-pass filters. The system decomposes an input image into a four-dimensional array based on spatial location ($x, y$), scale ($s$), and orientation ($o$).
By decomposing an image into 40 complex image pairs (5 scales $\times$ 8 orientations), the original image can be reconstructed by summing these components. This mathematical framework allows for "smoothly degrading quality" rather than sharp, artificial edges, ensuring that experimental results reflect missing information rather than processing noise.
Creating Complementary Stimuli
To generate complementary stimuli, the authors selectively mask specific Gabor coefficients. For instance, Image A might retain vertical information at a fine resolution, while Image B retains horizontal information at a coarse resolution. To maintain recognizability, "anchor points" (salient features like eyes or corners) are manually selected, and information is retained within windows around these points while being masked elsewhere.
This technique allows researchers to move beyond simple pixel-based differences. True complementarity is achieved in the "cortical representation" (the Gabor space), mimicking how the brain might partition visual information.
Results and Discussion
The authors successfully demonstrated that recognizable gray-scale images could be reconstructed from Gabor coefficients, despite minor quantization errors in extreme frequencies. The methodology provides a flexible tool for vision research, allowing precise control over the resolution, position, and orientation of realistic stimuli.
While the manual selection of anchor points and the computational intensity of convolution operations were limitations in 1995, the work laid the foundation for investigating how different orientation and scale "channels" in the brain interact during object recognition. This approach bridges the gap between simple line-drawing experiments and the study of complex, ecologically valid visual scenes.
Summary updated