This repository documents the synthetic image generation pipeline developed as part of a computer vision R&D project. The primary goal of this work is to generate realistic, annotated microscopy images by composing segmented blood cell instances onto clean backgrounds, enabling effective training and evaluation of object detection models.

⚠️ Scope note: This repository includes work completed up to April 2025. From Week 10 onwards, the focus shifts specifically to cell segmentation and synthetic image composition.

This repository focuses on:

• Segmentation of individual blood cells from microscopy images
• Synthetic image creation via controlled composition of segmented instances
• Automatic generation of bounding-box annotations in YOLO format
• Visual validation using object detection model outputs

Individual blood cells are segmented from microscopy images to create reusable foreground instances. The pipeline extracts clean cell instances across multiple classes for downstream composition.

Figure 1: Representative examples of different blood cell classes extracted from the base dataset. Each column shows instances of a specific cell type used as building blocks for synthetic image generation.

The segmentation pipeline uses classical image processing techniques, including:

• Color space transformations (HSV / LAB)
• Contrast enhancement using CLAHE (Contrast Limited Adaptive Histogram Equalization)
• Thresholding methods such as Otsu's thresholding
• Morphological operations (opening, closing, erosion, dilation)
• Watershed-based separation for touching cells

Segmented cell instances are randomly placed onto clean backgrounds to generate synthetic microscopy images. The composition process:

• Randomly samples cells from different classes
• Places them with controlled spacing to avoid excessive overlap
• Generates corresponding YOLO-format bounding box annotations
• Exports both the composite image and annotation file

Generated synthetic images are validated using a trained YOLO object detection model to verify the quality and realism of the composites.

Figure 2: Validation of synthetic composite images. Ground-truth bounding boxes (green) are automatically generated during composition, while YOLO model predictions (red) demonstrate successful detection of synthesized cells.

• Detailed Project Report – Complete methodology, experiments, and results

Work is organized chronologically into week-specific folders:

• Week1/ – Week9-10/: Foundational computer vision tasks and preliminary experiments
• Week10 onwards/: Cell segmentation pipeline and synthetic dataset generation
• Week11/, Week12/: Refinements, validation, and final integration

Each folder contains notebooks, scripts, and visual outputs relevant to that phase of development.