Microscopic Pollen Image Classification via Contour-Signal Representation, Wavelet Analysis, and CNN

Area of Science:

• Microscopy and Image Analysis
• Computational Biology
• Machine Learning

Background:

• Accurate classification of microscopic pollen grains is difficult due to image noise, variability, and complex backgrounds.
• Existing methods often struggle with weak textures and similarities within pollen classes.

Purpose of the Study:

• To develop a robust hybrid framework for microscopic pollen image recognition.
• To address challenges like noise, structural variability, and background complexity in pollen classification.

Main Methods:

• A hybrid framework integrating contour-signal modeling, spectral-wavelet analysis, and deep learning.
• Conversion of pollen images to contour-based point-signal representations.
• Application of Gaussian, median, and contour-aware filtering for signal enhancement.
• Spectral-wavelet analysis (Fourier, CWT, DWT) for feature extraction.
• Classification using a convolutional neural network (CNN).

Main Results:

• The proposed hybrid method significantly outperformed traditional computer-vision and baseline deep-learning approaches.
• Achieved an error rate of 6.4% with an overall classification accuracy of 0.977 and an F1-score of 0.966.
• Demonstrated noise-robustness and effectiveness in handling challenging microscopic image conditions.

Conclusions:

• Combining contour-based signal processing with deep feature learning offers an effective strategy for microscopic pollen image recognition.
• The framework shows promise for robust classification in challenging microscopic imaging scenarios.
• Further validation on diverse micro-object datasets is recommended to assess broader generalization.