jnoxro / joker Goto Github PK

code is inconsistent between train.cpp and joker.cpp, train.cpp should be updated to meet the same optimizations joker.cpp has received

Most of the code is missing comments or hard to understand.

Usage and installation information incomplete/missing

threads do the assignments contained within a job.

eg:

Job 1 - ocr
assignment 1 - compare target with model 1,2,3
assignment 1 - compare target with model 3,4,5
...
Job 2 - ocr
...
Job 3 - img threshold
...

Build system requires certain libs for GraphicsMagick build which may not be present / GraphicsMagick may already have been built without these. Checks should be made with cmake.

So threads can carry out many different tasks whilst program runs (eh threshold actions, OCR, training), they OCR controller should be a new class, with functions to change private variables threads can access for jobs

negative scoring values for training would currently be trail and error for users.

Training should test a range of negative values automatically to find optimal

Add building the jk executable as a cmake option.

If users only want to use the library there is no need for a local GM install, plus they may want extra GM features.

Let them hand it themselves.

using pointers to pass memory locations, +=, -+, and other direct modifications instead of copy operations will improve efficiency

Only operations carried out by GraphicsMagick are image read and threshold > could try read that directly and remove the dependency

Joker has only been tested (build and binary) on systems set up for development with multiple libraries and GraphicsMagick already installed. In order to make robust installation instructions tests need making on clean distros

Model save / load cuts out 0's, this can be used for any repeating (>2) values

the flags shown are not in order, they should be separated into optional and required, and have the default values shown for optional

Model format and save/load functions were left after a working version found.

I/O not saturated - 3ms to open 35.3kB model = 12MB/s so lots of room for improvement.

Init OCR should have another overload allowing for passing of PixelCache

when training a model pixelaverage simply sums how often a pixel contains data, this should be converted to fixed range (eg 0-100, or 0-1000) to prevent massive number calculations

Adding a new job to the pool will replace the OCR target even for threads still working on the previous job.

a well matching A may receive a massively different score than a well matching W for example. An 'ideal score' or similar should be found for each model character, which can then be used to have the score for each model char be a function of that models ideal score.

Thread job size currently set number (eg 2 comparisons) job size should be total jobs/threadcount

CMakelists.txt is hacked together to something that happens to build. This should be rewritten.

Some cases in which score should be reduced results in score increase eg:

Model = negative score Val, image = present
(-neg Val) * (-1 for present) = increase

Instead could use 0 for model negative scoring positions, then use an if statement to reduce score.

EVERY case of misalignment should decrease score.

currently the highest scoring result is output, no matter the score.

This may remain, but an ocr class function should be added to output the score so developers can set their own minimum scores.

unsure if its a good idea to keep these present - they are .files so are hidden once cloned