We tested super-resolution pre-filter for LPR OCR. It did nothing

#1

Post by xmichael909 » Wed, May 13, 2026, 4:12 AM UTC

We tested super-resolution pre-filter for LPR OCR. It did nothing
wink.co

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#2

Post by xmichael909 » Wed, May 13, 2026, 4:12 AM UTC

Quick context: SR-as-a-pre-filter for ALPR gets a lot of attention right now there's a whole ICPR 2026 competition track on low-res plate recognition; so we actually testing it on our production lpr data.

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#3

Post by mirsadm » Wed, May 13, 2026, 6:35 AM UTC

xmichael909 wrote: post #2

Quick context: SR-as-a-pre-filter for ALPR gets a lot of attention right now there's a whole ICPR 2026 competition track on low-res plate recognition; so we actually testing it on our production lpr data.

Have you tried classic CV SR algorithms (eg SR from multiple nearby frames)? Single frame SR seems unlikely to be useful.

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#4

Post by xnx » Wed, May 13, 2026, 10:34 AM UTC

Not one image on the page?

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#5

Post by xmichael909 » Wed, May 13, 2026, 1:39 PM UTC

xnx wrote: post #4

Not one image on the page?

Yah we should probably add some back, we had examples but they were mostly meaningless to the human eye, but we'll try adding some "examples"

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#6

Post by xmichael909 » Wed, May 13, 2026, 2:12 PM UTC

mirsadm wrote: post #3

xmichael909 wrote: post #2

Quick context: SR-as-a-pre-filter for ALPR gets a lot of attention right now there's a whole ICPR 2026 competition track on low-res plate recognition; so we actually testing it on our production lpr data.

Have you tried classic CV SR algorithms (eg SR from multiple nearby frames)? Single frame SR seems unlikely to be useful.

No, that is a really great point. The ICPR competition the big winners were all about multi-frame. Will do some more experiments and update.

Re: We tested super-resolution pre-filter for LPR OCR. It did nothing

#7

Post by xmichael909 » Wed, May 13, 2026, 6:35 PM UTC

mirsadm wrote: post #3

xmichael909 wrote: post #2

Quick context: SR-as-a-pre-filter for ALPR gets a lot of attention right now there's a whole ICPR 2026 competition track on low-res plate recognition; so we actually testing it on our production lpr data.

Have you tried classic CV SR algorithms (eg SR from multiple nearby frames)? Single frame SR seems unlikely to be useful.

We tried that, wasn't a big lift to setup and tested. We used 1000 detections with at least 4 crops, and it made a bit of a difference but not much - our conclusion remains the same.