Self-normalized density map (SNDM) for counting microbiological objects.

Area of Science:

• Computer Vision
• Machine Learning
• Microbiology

Background:

• Density map (DM) approaches are used for counting microbiological objects in images.
• Existing DM models, like U-Net, have limitations in accurately predicting object counts due to statistical uncertainties.

Purpose of the Study:

• To analyze the statistical properties and uncertainties of the DM approach for object counting.
• To propose an improved DM model that addresses the deficiencies of existing methods.

Main Methods:

• Utilized bootstrap and Monte Carlo (MC) dropout for statistical analysis of deep neural networks.
• Developed a self-normalization module integrated into the U-Net architecture, creating the Self-Normalized Density Map (SNDM) model.

Main Results:

• The Self-Normalized Density Map (SNDM) model demonstrates improved accuracy in predicting the total number of objects.
• SNDM exhibits consistent statistical results across bootstrap and MC dropout frameworks, unlike the original model.
• The efficiency of SNDM is comparable to established detector-based models like Faster R-CNN and Cascade R-CNN.

Conclusions:

• The proposed self-normalization module significantly enhances the accuracy and reliability of density map-based object counting.
• SNDM offers a robust and efficient alternative for microbiological object detection and counting in images.