1

Topic: and preparation of sampling for training

Good afternoon, training  for recognition of hand-written digits needs many pictures of each digit. If I want, that the digit was recognized turned on 90 degrees appropriate pictures are necessary at training. An operation principle if I correctly understand such that in network layers the coefficients, tutorings satisfying to sampling steal up. I.e. as though  descent? Only it is not clear as to be, if it is necessary to learn to recognize digit on the big pictures where the digit can be at any angle and with any geometrical distortions (within recognition ). It is necessary for digit training under each corner to submit? And how  if for one digit at training we submit tens , and it is necessary to search for inputs in the big picture on thousand ?  I Am sorry, if a question silly.

2

Re: and preparation of sampling for training

Hello, Ocenochka, you wrote: O> And how  if for one digit at training we submit tens , O> and it is necessary to search for inputs in the big picture on thousand ? If to recognize only pictures with digits (and it is more than anything on them is not present) pictures  under one size, it is normal no more 100100 pixels. Yes, one of the main problems  - enough great volume of the data is necessary for training. Rather often given generate program. In case of digits them it is easily possible to distort-blur-peervorachivat, generate the new. But generally is a lot of already ready  on various subjects. It is necessary to look for them before to create the. To search of digits on the big images apply CNN networks, for which too all images  in a uniform format. Instead of  how work CNN networks, I will recommend courses on . Here rather basic, simple and intelligible from Adrew NG (a specialization part )

3

Re: and preparation of sampling for training

Hello, Ocenochka, you wrote: O> training  for recognition of hand-written digits O> needs many pictures of each digit. O> if I want, that the digit was recognized turned on 90 degrees, O> that appropriate pictures are necessary at training. O> an operation principle if I correctly understand such that O> in network layers the coefficients, tutorings satisfying to sampling steal up. O> I.e. as though  descent? O> only it is not clear as to be, if it is necessary to learn to recognize digit on the big pictures, O> where the digit can be at any angle and with any geometrical distortions (within recognition ). O> It is necessary for digit training under each corner to submit? O> and how  if for one digit at training we submit tens , O> and it is necessary to search for inputs in the big picture on thousand ? O>  I Am sorry, if a question silly. You take mnist, you take Caffe. At training it is recommended to do for pictures  distortions, including turns. You use convolution networks. The specific public project is here: textdet.com - there all is simple enough, though accuracy not the super.