Neural

If one has an input video and one frame of it processed to get a desired image ( noise free or some other feature) then Neural network can be trained on that particular input frame and output target frame to get internal processing to match. For example if a 5x5 grid is used on that particular frame (with unknown weights) to get the target output then Neural learns by training how to get the nearest output. Works only if the target frame is result of linear processing of input

This Neural function assumes that the first frame of the input video was processed to get the tfn frame of the training clip. A window on the frame (that has all the range of amplitudes) is used for training. Starting with random weights for each input nodes through number of iterations the weights are adjusted with Resilient algorithm. This training may take several seconds to a few minutes.

After training these weights are used to process the entire video including the first frame. In case of RGB input, which color component be used for training can be specified. For YUV formats training is done on Y component only. While all 3 components of RGB formats are processed , in case of YUV only Y is processed

In some situations the error during each iteration may reach a local minimum and diverge. Best weights are stored for usage. In rare cases starting weights may not be suitable, in which case one can choose trying with several different starting weights and use those that give least error. Also instead of the inbuilt random numbers, a different seed can be used. However it should be borne in mind that the tme for learning will increase. Time is proportional to xpts * ypts * window_size * iter * bestof

Training window needs to be as small as possible to minimise time, but needs to encompass all possible range of values. A minimum 100x100 is accepted. If target frame was processed with an x * y grid, then xpts must be a minimum x and ypts be a minimum y

Packed RGB ( 8 and 16 bit), YUY2 and 8, 10 to 16 or 32 bit Planar formats are supported. This is thread safe. MT_NICE_FILTER

To save time in future for same situation the derived weights can be saved to a file. Instead of spending time on training the saved weights can be read from file and used. However please be sure that the saved weights are valid in the present situation. This file also includes number of target points, error at each iteration for each weight set. If reading from previously saved file, parameters listed after fname need not be given.

Neural(a,b, iter = 3000 , wset = false, bestof = 1,tty = 100, tlx = 100, trx = 300, tby = 300, xpts = 7)
# another though trivial example, takes long time to train:-
a=colorbars()
b = GBlur(a,rad = 7, sd = 2.0)
Neural(a,b, iter = 1000, bestof = 6, wset = false,tty = 100, tlx = 200, trx = 400, tby = 300, xpts = 7, txt = "save", fname = "C:\temp\neuralfile.txt")