QRCodes decoding without Quirc

[dkurt]

Describe the feature and motivation

Currently, OpenCV is able to detect and encode QR codes with objdetect module. However decoding requires to build with Quirc library.

Using https://github.com/yansikeim/QR-Code/blob/master/ISO%20IEC%2018004%202015%20Standard.pdf (section 11: Decoding procedure overview) implement a decoding without quirc

Current progress on https://github.com/dkurt/opencv/tree/qrcode_decode branch:

• Version 1 with no error correction
• Error correction levels (Version 1)
• Consider alignment markers
• Performance analysis and improvements
• Check related issues:
  • QR code detecting in python with OpenCV raises UnicodeDecodeError: 'utf-8' codec can't decode byte #23728
  • Unable to decode QR Codes with particular payloads #23810

Additional context

No response

[AleksandrPanov]

@dkurt, there are no problems with building quirc library. Quirc library is built automatically from the 3rdparty folder.
The most important part of decoding is performed using OpenCV at the straightDecodingProcess() stage. The alignment markers are considered in detectAlignment() stage. Most of the problems are at the samplingForVersion() stage. At this stage, the QR code binarization is too simple.

bool QRDecode::straightDecodingProcess()
{
    if (!updatePerspective(getHomography()))  { return false; }
    if (!versionDefinition())  { return false; }
    if (useAlignmentMarkers)
        detectAlignment();
    if (!samplingForVersion()) { return false; }
    if (!decodingProcess())    { return false; }
    return true;
}

Please consider improving a specific stage or stages, instead of redoing the entire algorithm.

[dkurt]

@AleksandrPanov, going to remove quirc dependency at all. Or at least propose an alternative for decoding stage only (decodingProcess). Other stages remain unchanged.

Considering alignment markers during decoding is just an internal feature required for binary blocks extraction, not detection.

I can propose a minimal working version soon at least for version 1 codes to demonstrate the expectations.

[AleksandrPanov]

@dkurt:

1. I don't understand why to remove the quirc dependency if everything works.
2. Binary blocks are actually extracted at the samplingForVersion() stage. The decodingProcess() stage uses already extracted binary blocks and does not change them.
3. I propose to improve the extraction of binary blocks at the samplingForVersion() stage.

[AleksandrPanov]

The samplingForVersion() stage extracts binary blocks from the image. The stage returns: straight = Mat(Size(version_size, version_size), CV_8UC1)

We will not be able to effectively correct the binary blocks due to the reduced resolution.

[dkurt]

1. I don't understand why to remove the quirc dependency if everything works.

This is still a 3rdpatry dependency which should be maintained. There was a related GSoC task definition but there is no other motivations except avoid the dependency: https://github.com/opencv/opencv/wiki/GSoC_2020#idea-implementation-of-qr-codes-decoder-and-encoder (#17889)

2. Binary blocks are actually extracted at the samplingForVersion() stage. The decodingProcess() stage uses already extracted binary blocks and does not change them.

samplingForVersion returns a straight image with 0/1 modules, but as a part of decodingProcess there is a procedure of extracting symbol blocks by 8 modules. So for now quirc takes a bitstream from the straight QR code and decodes the data. Removing Quirc dependency means rewrite the decoding of the bitstream only.