Does the file fit?
Dimensions, format, aspect, and basic image constraints catch cheap failures before model work begins.
fast checksLive TUI demo · 0:17
Recorded run envelope. The ranges below are observed process measurements; the false-positive figure is a target threshold, kept visually separate from measured behavior.
95–98%text detection range
180–350msper-image processing
<3%false-positive target
8–12%manual-review band
01 · Architecture
Four gates, one honest output contract.
PhotoValidator uses a staged cascade with a parallel middle. Specification and border checks can short-circuit obvious failures; candidates that pass those cheap gates fan out to text, watermark, and perceptual-quality modules that can run concurrently before the confidence engine returns valid, invalid, or manual review.
Is the frame acceptable?
Canny edges, morphology, and contour reasoning identify border and framing violations.
OpenCV geometryWhat is visible?
ConvNeXt and PaddleOCR detect unwanted text and watermark patterns, with outputs kept separate.
parallel vision + OCRIs it usable?
BRISQUE, NIQE, and CLIPIQA combine into an adaptive quality decision with review when confidence is low.
parallel quality gate
02 · Vision, made visible
Every check you can see.
The documentation treats each processing stage as an inspectable figure. The same test image travels from the original frame through grayscale, Sobel gradients, and Canny edges; worked border, text, and watermark examples then carry their confidence scores and final classification on the page. Click any figure to read it at full resolution.
03 · Training and failure
470,000 synthetic images taught the wrong lesson.
The first training path produced a large synthetic dataset, but the result overfit the generation process instead of generalizing to the faculty’s visual conditions. The valuable engineering move was recognizing the failure and changing the pipeline, rather than treating dataset size as progress.
470ksynthetic-image pivot
The experiment exposed a distribution problem. The final system leans on staged checks, targeted detectors, shared preprocessing, caching, and a review band instead of trusting one overfit classifier.
04 · Engineering judgment
The important decisions protect people and operators.
Patch the OCR boundary.
PaddleOCR wrote outputs in a path shape that conflicted with the rest of the pipeline. A runtime monkey patch normalized that boundary without rewriting the model.
Do not crop away identity.
A face/shoulder crop was abandoned because it could cut religious headwear and natural features. Validation should respect the person in the image.
Make review explicit.
Manual review is a first-class destination. The system can load once, process in parallel, fall back to CPU or heuristics, and leave a clear queue when confidence is not enough.
Operator surface
The TUI exposes batch queues, stage-level counters, flagged files, and a review path rather than hiding the pipeline behind a single progress bar.
queued → spec → border → text → quality
valid · invalid · manual reviewMeasured envelope
Recorded behavior includes 98–99% specification checks, 90–95% border detection, 93–96% watermark detection, 85–92% quality scoring, 200–400 images per minute, and a 5–10 second cold start.
models loaded once · shared preprocessing · graceful fallback05 · Delivery
Production readiness is part of the case study.
The implementation was delivered with model memory controls, cache-aware execution, parallel processing, CPU/GPU fallback, reports, and the certificate above. The system is valuable because another operator can understand what it decided and what still needs attention.
06 · Complete record
Open the report when you want the whole build.
The page gives the visual and engineering story. The original local artifacts preserve the full documentation, certificate, and working demo.