RGB-Based Computer Vision and Spectral Characterization of Semiconductor Samples

RGB-Based Computer Vision and Spectral Characterization of Semiconductor Samples ## Project Overview I am developing a Python pipeline to investigate whether standard RGB photographs can predict useful spectral properties of semiconductor samples. My dataset contains approximately 62 RGB images of sample plates, with repeated photographs of some physical plates, and associated diffuse reflectance spectra. Each plate contains 48 samples. Correctly mapping images, sample IDs, and reference spectra is essential. ## Scope of Work **1. Image Processing and ROI Extraction** - Detect and crop plates, preserving existing working code wherever possible. - Reliably extract 48 sample regions and handle both 6 × 8 and 8 × 6 layouts. - Improve orientation detection and consistent sample labelling. - Test robustness across multiple images and imaging conditions. **2. Dataset Preparation** - Match physical plates, sample IDs, RGB images, and reference spectra. - Extract RGB features and prepare a traceable dataset. - Flag uncertain sample-to-spectrum correspondences rather than guessing. **3. Machine Learning and Evaluation** - Establish a simple baseline model before attempting deep learning. - Predict a scientifically justified target derived from the available spectra. - Validate performance without leakage from repeated photographs. - Report appropriate metrics, limitations, and whether RGB data support the intended prediction task. ## Deliverables - Complete, runnable Python source code and setup instructions. - Reproducible image-processing and prediction pipeline. - Visualizations of plate detection and ROI extraction. - Validation report and recommendations for future improvements. Pixel-level bandgap mapping and guaranteed prediction accuracy are not assumed deliverables; feasibility must be established from the available data. ## Milestones and Budget **Maximum total budget: €200.** Proposed allocations total €140, with €60 reserved for additional work subject to prior agreement. 1. **ROI extraction and orientation — €40:** Preserve working plate detection and cropping; improve grid detection, orientation, and sample labelling. 2. **Dataset and spectral alignment — €30:** Validate sample IDs and reference-spectrum correspondence. 3. **Baseline ML model — €50:** Train and evaluate an appropriate baseline using leakage-resistant validation. 4. **Documentation and handover — €20:** Deliver source code, instructions, and evaluation report. Each milestone must have agreed acceptance criteria before work begins. Existing working components must not be replaced without justification and approval. Any additional work requires prior agreement and must remain within the €200 total budget cap.