Mach edges: local features predicted by 3rd derivative spatial filtering.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Image Processing

Background:

• Edges are crucial for visual scene interpretation.
• Standard edge detection models often rely on luminance gradient maxima or second derivative zero-crossings (ZCs).
• These models assume edges correspond to the steepest luminance changes.

Purpose of the Study:

• To investigate human edge perception using a novel stimulus.
• To test the validity of standard edge detection models against human perception.
• To explore alternative models for edge detection.

Main Methods:

• A stimulus with a linearly ramped luminance gradient (blurred triangle wave) was used, lacking gradient maxima and ZCs.
• Human observers marked perceived edge locations on this stimulus.
• The locations of perceived edges were analyzed in relation to derivative operators of the luminance profile.

Main Results:

• Observers consistently identified edges at the corners of the luminance gradient profile.
• These perceived edges did not correspond to gradient maxima or ZCs.
• The identified Mach edges aligned with peaks and troughs in the third derivative of the luminance profile.
• A recent two-stage model involving derivative operators and rectification accurately predicted these Mach edges.

Conclusions:

• Human edge perception is not solely based on gradient maxima or second derivative zero-crossings.
• Mach bands, perceived at gradient corners, challenge conventional edge detection algorithms.
• A recent computational model, utilizing a sequence of first and second derivative operators with rectification, provides a better account of human Mach edge perception.