The April 2026 check in contrast PP-OCRv5, GLM-OCR, Tesseract and EasyOCR on cropped handwritten prescription phrases. GLM-OCR produced the very best character-level efficiency, recording a Character Error Charge, or CER, of 0.328. PP-OCRv5 delivered the decrease Phrase Error Charge, or WER, at 0.789, in contrast with 0.801 for GLM-OCR. Decrease scores point out fewer errors below each measures.
Precise-word recognition confirmed a a lot sharper divide between the newer and older applied sciences. GLM-OCR appropriately reproduced 32.6% of examined phrases, whereas PP-OCRv5 achieved 21.4%. Tesseract managed 2.5% and EasyOCR 2.6%, giving the 2 newer methods roughly an eightfold to 13-fold benefit relying on the comparability.
The outcomes underline an vital shift in OCR improvement. PP-OCRv5 makes use of solely about 5 million parameters, but its word-level efficiency remained aggressive with GLM-OCR, a 900-million-parameter multimodal mannequin. Analysis behind PP-OCRv5 argues that fastidiously chosen, precisely labelled and numerous coaching information can permit specialised light-weight OCR methods to compete with far bigger vision-language architectures.
GLM-OCR represents the choice method. Its structure combines a roughly 400-million-parameter visible encoder with a 500-million-parameter language decoder. The mannequin is designed to interpret doc photos whereas utilizing linguistic context to enhance transcription. Its multi-token prediction mechanism additionally permits a number of tokens to be generated throughout a decoding step, lowering among the pace penalty usually related to generative vision-language methods.
That language functionality seems significantly helpful when particular person letters are troublesome to differentiate. Within the prescription benchmark, GLM-OCR’s CER was significantly higher than PP-OCRv5’s 0.477. But PP-OCRv5 carried out marginally higher on WER, suggesting that stronger character-level reconstruction doesn’t mechanically translate into fewer complete-word errors.
Generative OCR additionally introduces totally different failure modes. Throughout testing, GLM-OCR generally produced formatting or characters that weren’t current within the handwritten enter. One four-letter pattern was interpreted as mathematical notation quite than strange textual content. PP-OCRv5, primarily based on a extra standard specialised recognition pipeline, was much less liable to this kind of generative deviation, though it nonetheless misinterpret giant numbers of adverse phrases.
The experiment used RxHandBD, a prescription handwriting dataset containing 5,578 cropped phrase photos and 1,559 distinctive textual content entries. The fabric consists of generic medication names, pharmaceutical manufacturers, dosage kinds and scientific terminology drawn from bodily prescriptions. A ready model divides the dataset into 4,463 coaching photos and 1,115 testing photos.
These traits additionally restrict how broadly the findings may be utilized. The benchmark used cropped particular person phrases quite than full prescription pages, that means it didn’t absolutely check the problem of finding handwriting, separating printed materials and preserving relationships between drug names, doses and directions. It was additionally a single CPU-only run utilizing default mannequin configurations with out medical fine-tuning.
Handwritten prescription recognition stays troublesome as a result of medical paperwork mix extremely variable writing, abbreviations, unfamiliar drug terminology and contextual data that may alter the that means of in any other case comparable textual content. Analysis into multimodal prescription interpretation has highlighted the extra issue created by restricted datasets and the lack of contextual data when prescriptions are divided into remoted phrase photos.
The scientific threshold is significantly larger than the benchmark outcomes. Even GLM-OCR’s 32.6% exact-match rating means roughly two phrases out of three weren’t reproduced completely. A system making errors at that charge can’t safely decide medicines, strengths or dosing directions with out human verification. The benchmark itself was an impartial technical check quite than a scientific validation examine and didn’t consider affected person outcomes or meting out security.
