Baidu's 6 MB OCR Model Runs Entirely in the Browser with No Backend

Client-side OCR at 6 MB removes the operational overhead of deploying, scaling, and paying for a Python inference service. For internal tools and low-volume apps, the browser handles the GPU work directly, and the model is small enough to load over a spotty connection without a noticeable wait.

Baidu's PP-OCRv6 tiny packs a text-detection model (1.74 MB) and a text-recognition model (4.28 MB) into a combined 6 MB payload that loads in the browser and runs inference through ONNX Runtime Web on WebGL or WebGPU. The pipeline resizes images, runs DBNet-based detection with binarization and connected-component post-processing, then feeds cropped text regions into a CRNN recognizer with CTC greedy decoding. The entire flow—upload, detect, recognize, display—happens client-side with no backend service.

Three concrete pitfalls surfaced during the build. A mismatched character dictionary caused garbled output because the ONNX model outputs 6,906 classes but the downloaded dictionary held only 6,622 entries; extracting the character set directly from the model's protobuf metadata fixed it. NaN values in model output propagated through softmax and required explicit finite-value guards during argmax and exponentiation. A 0.3 confidence threshold on softmax over 6,906 dimensions wiped out all results because per-character probabilities rarely exceed 0.05.

Switching to the small model variant broke the frontend post-processing entirely: small uses a single-channel binary map with contour detection instead of DBNet's three-channel probability map and unclip expansion, so the detection pipeline needs a rewrite.