Deep Learning and Histogram-Based Grain Size Analysis of Images

Area of Science:

• Geosciences
• Sedimentology
• Computational methods

Background:

• Grain size analysis is crucial for understanding sedimentary environments and hydrodynamic conditions in sedimentary simulation experiments (SSEs).
• Traditional image-based grain size analysis methods struggle with the fuzzy grain edges and irregular arrangements common in SSE images.

Purpose of the Study:

• To develop a deep learning model for accurate grain size analysis of images from SSEs, addressing limitations of existing methods.
• To enhance the quantification and automation of grain size analysis in SSEs.

Main Methods:

• A deep learning model combining ResNet18 feature extraction with histogram layers was proposed.
• Local histogram features were extracted and concatenated to form comprehensive image histogram features.
• These features were used to estimate grain size corresponding to cumulative volume percentage.

Main Results:

• The proposed deep learning method achieved higher accuracy compared to eight other models.
• Results showed high consistency with manual grain size analysis.
• The model effectively handles images with irregular grain distribution and fuzzy edges.

Conclusions:

• The developed method significantly enhances the efficiency and accuracy of grain size analysis in SSEs.
• This approach improves the quantification and automation of grain size analysis for complex sedimentary images.
• The method has potential applications in soil science and geotechnical engineering.