Silicrop

A PyQt5-based application for silicon wafer image processing, featuring AI-powered ellipse detection and correction.

Overview

Silicrop is designed to process images of silicon wafers by:

• Detecting wafer boundaries using deep learning (DeepLabV3 segmentation)
• Fitting ellipses to wafer contours
• Correcting perspective distortion to produce circular images
• Detecting and handling "meplat" (flat edge) regions on 150mm wafers
• Supporting batch processing for multiple images

Features

• AI-Powered Detection: Uses a DeepLabV3 model with ResNet-18 backbone for wafer segmentation
• Interactive GUI: PyQt5 interface with drag-and-drop image loading
• Manual Adjustment: Click to add/move/remove ellipse fitting points
• Notch Detection: Mark notch regions with 3-point triangles (press 'N' to toggle mode)
• Batch Processing: Process multiple images in parallel
• Mask Export: Save segmentation masks for training data

Installation

Requirements

• Python 3.8+
• PyQt5
• PyTorch
• OpenCV (cv2)
• NumPy
• Pillow
• segmentation-models-pytorch
• matplotlib
• scipy
• joblib
• psutil

Install

pip install -e .

Or install dependencies directly:

pip install PyQt5 torch torchvision opencv-python numpy pillow segmentation-models-pytorch matplotlib scipy joblib psutil

Usage

Launch the Application

python -m silicrop.main

Or if installed:

silicrop

Basic Workflow

1. Select a Model: Click "Select a model" to load a trained .pth model file
2. Load Images: Drag and drop images into the left panel
3. Select Scale Mode: Choose "<=150" for wafers with flat edges, ">200" for full circular wafers
4. Process:
   • Single image: Select image and click "Run AI"
   • Batch: Click "Select All" then "Run AI (batch)"
5. Save Results: Use "Save Image" or "Save Mask" buttons

Keyboard Shortcuts

• N: Toggle notch drawing mode (for marking notch regions)
• Ctrl + Drag: Pan the image view
• Scroll: Zoom in/out

Project Structure

silicrop/
├── __init__.py          # Package version
├── main.py              # Application entry point
├── assets/
│   └── icon.png         # Application icon
├── layout/
│   ├── components.py    # Main GUI (ImageProcessorApp)
│   ├── dragdrop.py      # File drop list widget
│   ├── how_to.py        # Help popup
│   ├── styles.py        # UI stylesheets
│   └── theme.py         # Application theme
└── processing/
    ├── crop.py          # Ellipse fitting (FitAndCrop)
    ├── meplat.py        # Flat edge detection
    ├── prediction.py    # AI model (EllipsePredictor)
    ├── prediction_batch.py  # Batch processing
    ├── rotate.py        # Image rotation (Rotate)
    └── utils.py         # Mouse navigation handler

Key Classes

ImageProcessorApp

Main application window containing:

• File list with drag-and-drop
• Original image viewer (FitAndCrop widget)
• Processed image viewer (Rotate widget)
• Control buttons for processing and saving

EllipsePredictor

AI inference engine that:

• Loads DeepLabV3 segmentation model
• Predicts wafer masks from images
• Extracts contours and fits ellipses
• Handles meplat detection for 150mm wafers

FitAndCrop

Interactive widget for:

• Displaying original images
• Manual 5-point ellipse fitting
• Notch region marking (3 points)
• Perspective transformation to circular output

Rotate

Widget for:

• Displaying processed (circular) images
• Interactive rotation by clicking
• Line-based rotation (150mm mode)
• Point-to-bottom rotation (200mm mode)

Model Training

The application expects a trained DeepLabV3 model. Training code is available in training/model.py (not used at runtime).

Model Architecture

• Encoder: ResNet-18 (pretrained on ImageNet)
• Decoder: DeepLabV3
• Output: Binary segmentation mask