Lesson 2. Image → tensor — Train CNN and CRNN from scratch
The series Train CNN and CRNN from scratch: cell crop → tensor [1, 48, 160]. Gray channel, safe contrast, letterbox without distortion, and inversion — before any model.
Stuzhuk Lab — Chemistry of Code
Chemistry of Code
← All posts
Tag
computer vision
All blog posts with this tag.
Computer Vision without a framework lock-in: 15 principles for PyTorch and TensorFlow
Computer Vision is not a PyTorch-vs-TensorFlow debate. Data, tensors, CNNs, classification, detection, segmentation, transfer learning, metrics, and production form one pipeline that survives an API change.
ONNX: from PyTorch to Runtime — format, export, and inference parity
What ONNX and ONNX Runtime are, how to export from PyTorch, opsets and dynamic axes, why CRNN/YOLO exports break, and how to match Python before the browser.
WebGPU + Transformers.js: neural nets straight in the browser
Run ML in React without a mandatory server GPU: WebGPU, Transformers.js, ONNX Runtime, OCR and computer vision, WASM fallback, and model-size budgets.
YOLO: how a network learns to find objects in an image
Object detection in plain engineering language: how YOLO differs from classification and segmentation, what a box, confidence, and duplicate suppression actually mean, why pretrained weights and your own dataset matter, and how to connect a detector with OCR, lines, a graph, and a language model on process diagrams.
Cropping a UT thickness form: projections, gated perspective, and other sheet-splitting methods
Field notes: how the pilot cuts a UT form photo into cells — morphology, projections, gated perspective — and which other sheet-splitting methods show up in industry practice.
Neural network quality metrics: how to know the model actually works
How to choose neural network quality metrics: classification, computer vision, OCR, LLM and RAG, calibration and production indicators — what to measure so the model truly works in the product.
A compact CRNN for UT measurements: 640k parameters and empty cells
Field notes: a local CRNN on thickness crops, CTC, 640k parameters, the empty-cell conflict, and why a Transformer is not the first step.
Empty cells, dittos, and strike-through: OCR should not decide if a measurement exists
UT pilot sequel: five cell kinds, ink, ditto-up-the-column, strike-through, and why heuristics still do not replace a tiny classifier.
AI analysis of weld radiographs: from dataset to inspection system
AI for weld radiography: dataset, labels, detector, rules, and an expert. Why 88 GDXray images are not a plant dataset.
Handwritten UT forms: why one model is not enough
Field notes from a pilot: computer vision, OCR, vision LLMs, confidence, and operator review before ERP — not a finished production guide.
Teaching a Neural Network to Read Handwritten Digits: From CNN to TrOCR
A practical path from single-digit classification to sequence recognition on real paper forms: CNN, CTC, CRNN, TrOCR, labels from your database, experiments, and a production pipeline without the wrong model.