Batch Image Enhancement
Automated image quality improvement at scale: brightness/contrast normalisation (CLAHE), noise reduction (Gaussian blur, bilateral filter, deep learning denoising), sharpening, colour correction, format standardisation (resize, crop, WebP/AVIF conversion). Libraries: OpenCV, Pillow. Throughput: 1,000-10,000 images/minute on CPU, 100,000+ on GPU.
Background Removal & Product Images
AI-powered background removal for e-commerce product images: semantic segmentation to isolate product from background (U2-Net or SAM), clean white or transparent background, shadow addition/removal, consistent aspect ratio padding. rembg (U2-Net) production-ready. Batch S3 pipeline with e-commerce platform integration (Shopify, WooCommerce via webhook).
OCR & Document Image Processing
Optical character recognition pipelines for scanned documents. Pre-processing: deskewing, denoising, binarisation, contrast enhancement. OCR: Tesseract 5 (open-source, self-hostable), AWS Textract (layout-preserving structured extraction), Google Document AI (pre-built invoice/form models). Post-processing: spell correction, confidence filtering, format reconstruction. Handles multi-column layouts, tables, handwriting.
Image Quality Assessment (IQA)
Automated filtering of low-quality images before they reach production or annotation pipelines: BRISQUE (no-reference quality score, fast), NIQE, focus detection (Laplacian variance identify blurry images), overexposure/underexposure detection, ML-based IQA (MUSIQ, NIMA). Use cases: filter blurry product images, prioritise manual review queue by quality, reject uploads below threshold in real time.
Medical Image Processing (DICOM)
DICOM (Digital Imaging and Communications in Medicine) pipeline development: DICOM file parsing and metadata extraction (pydicom), pixel data conversion (windowing map raw DICOM values to 8-bit display values per modality), series reconstruction (3D volume from CT/MRI slice stacks SimpleITK), anonymisation (remove patient identifiers for research), format conversion (DICOM to PNG/JPEG for ML training), pre-processing for ML (MONAI framework the medical imaging equivalent of torchvision).