Environmental Microorganism Image Analysis Using Deep Learning

A group applied AI project using the EMDS dataset to compare end-to-end image-analysis methods from preprocessing through localization.

Problem / goal

Microorganism image analysis requires consistent preprocessing and model evaluation across segmentation, classification, and localization tasks that are often treated separately.

My role

Applied AI Developer on a group deep learning project.

What I personally contributed

• Compared DnCNN and TROF denoising with U-Net, U-Net++, and DeepLabV3+ segmentation pipelines.
• Benchmarked handcrafted-feature, CNN, and Transformer classifiers and evaluated Faster R-CNN for object localization.
• Measured how preprocessing choices affected downstream IoU, Dice, and classification F1-score.

Technical approach

• Structured denoising, segmentation, classification, and localization as modular experiments with shared evaluation outputs.
• Compared CNN, Transformer, and handcrafted-feature approaches using task-appropriate metrics.
• Connected preprocessing quality to downstream segmentation and classification behavior through controlled comparisons.

Key features

• Denoising, segmentation, classification, and localization stages.
• CNN, Transformer, and handcrafted-feature comparisons.
• Modular experiment workflow for maintainability.
• IoU, Dice, and F1-based evaluation.

Impact / results

• U-Net++ after DnCNN preprocessing achieved 0.9365 IoU and 0.9843 Dice.
• ResNet18 achieved a 0.9464 F1-score for microorganism classification, with Faster R-CNN used for localization experiments.

What I learned

• Preprocessing quality can materially change downstream segmentation performance.
• A modular pipeline makes multi-stage computer vision experiments easier to compare, debug, and reproduce.