Environmental Microorganism Image Analysis Using Deep Learning

A group deep learning project that explored a full environmental microorganism image-analysis pipeline from preprocessing through detection.

Problem / goal

Manual preprocessing and model selection can slow microorganism image analysis, especially when denoising, segmentation, classification, and detection are treated as disconnected tasks.

My role

Collaborator on a group deep learning project.

What I personally contributed

• Compared denoising, segmentation, classification, and object detection methods in a modular image-analysis pipeline.
• Evaluated segmentation and classification with task-appropriate metrics including IoU, Dice, and F1-score.
• Connected preprocessing quality to downstream model performance through comparative experiments.

Technical approach

• Compared TROF and DnCNN denoising, U-Net variants and DeepLabV3+ segmentation, handcrafted-feature SVMs, CNN/Transformer classifiers, and Faster R-CNN detection.
• Used a modular pipeline design to reduce manual preprocessing and make model comparisons easier to maintain.
• Evaluated segmentation and classification using task-appropriate metrics including IoU, Dice, and F1-score.

Key features

• Denoising, segmentation, classification, and detection stages.
• Comparison of CNN, Transformer, and handcrafted-feature methods.
• Modular pipeline design for maintainability.
• Evaluation across multiple image-analysis tasks.

Impact / results

• U-Net++ achieved 0.9365 IoU and 0.9843 Dice after DnCNN preprocessing.
• ResNet18 achieved 0.9464 F1-score for microorganism classification.

What I learned

• Preprocessing quality can materially affect downstream segmentation performance.
• Pipeline modularity makes computer vision experiments easier to debug and compare.