A relaxation method for multispectral pixel classification

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Multispectral pixel classification is crucial for image analysis.
• Reducing classification errors is essential for accurate image interpretation.
• Existing methods like postprocessing and preprocessing have limitations.

Purpose of the Study:

• To compare the effectiveness of three distinct approaches for reducing errors in multispectral pixel classification.
• To evaluate postprocessing, preprocessing, and relaxation techniques.
• To determine the superior method for enhancing classification accuracy.

Main Methods:

• Postprocessing: Iterated reclassification based on neighbor class comparison.
• Preprocessing: Iterated smoothing via averaging with selected neighbors before classification.
• Relaxation: Probabilistic classification followed by iterative probability adjustment.

Main Results:

• The relaxation approach demonstrated markedly superior performance in experiments.
• Relaxation eliminated 4-8 times more classification errors compared to postprocessing and preprocessing.
• A color image of a house was used as a test case.

Conclusions:

• The relaxation method is highly effective for reducing errors in multispectral pixel classification.
• Probabilistic classification with iterative adjustment offers significant advantages over traditional methods.
• This research highlights relaxation as a key technique for improving image classification accuracy.