Color and luminance processing in V1 complex cells and artificial neural networks

Area of Science:

• Computer Vision
• Neuroscience
• Artificial Intelligence

Background:

• Object recognition relies on identifying object boundaries.
• Superposition of luminance and color edges is a key cue for boundary detection.
• Convolutional neural networks (CNNs) are models for image recognition.

Purpose of the Study:

• To investigate the role of color and luminance edge detection in CNNs for object recognition.
• To understand how CNNs process visual information for identifying object boundaries.
• To compare CNN visual processing with biological visual systems like V1 complex cells.

Main Methods:

• Examined CNN models trained on natural images.
• Focused on units in the second convolutional layer.
• Analyzed unit activations for invariance to spatial phase and tuning for color and luminance.

Main Results:

• Some units showed tuning for a nonlinear combination of color and luminance.
• Other units were tuned for luminance alone.
• Few units were tuned for color alone, similar to V1 complex cells.

Conclusions:

• CNNs exhibit sensitivity to combined color and luminance, supporting object boundary detection.
• This pattern suggests an efficient visual recognition strategy, potentially robust to lighting variations.
• The lack of color-alone invariance implies redundancy with other visual representations.