updates

Author: matrn

README.md

@@ -17,6 +17,27 @@ Handwritten digits recognition in C using neural network trained using [MNIST da
 	- use `make clean && make release` to generate AppImage binary (you have to install `linuxdeploy` and other dependecies using `make install_tools` first)
 
 
+# GUI
+![GUI](./data/assets/C-digit-recognition-example-of-usage.jpg)
+
+## Control
+- `left mouse button & drag` to draw
+- `right mouse button` to clear the draw space
+- `middle mouse button or Recognise button in the GUI` to run recognition process
+
+## Processing of the drawn image
+Preprocessing used in [MNIST](http://yann.lecun.com/exdb/mnist/) database: *The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and translating the image so as to position this point at the center of the 28x28 field.*
+
+1. **crop calculation** - crop of the whole draw space is calculated from sides until it reaches non white pixels. After that maximum of the width & height of the cropped image is then taken and `crop_x` & `crop_y` & `crop_w` & `crop_h` is recalculated to preserve image ratio.
+2. **sub image generation** - using previous values and drawn image stored in pixbuf sub image is created
+3. **scaling** - previous cropped image is scaled to the `20x20` image
+4. **conversion into grayscale** - pixbuf is converted into `uint8_t` grayscale image
+5. **adding frame** - `4, 4, 4, 4` frame is added to the `20x20` image resulting into `28x28` image
+6. **computation of the center of mass of the pixels** - is done using mean values accross X & Y
+7. **move of the submatrix** - submatrix (drawn number) is moved in the framed image
+8. **neural network forward propagation** - this preprocessed image is fed to the neural network
+
+
 # Libraries
 Dependency of the libraries is in this order: `GUI -> ceural -> lag`. For documentation see source code or use IDE (for example vscode).
 
@@ -36,25 +57,7 @@ Library supports many operations but more development is needed because currentl
 ## Ceural
 Ceural library is created for multi-layer networks trained using MNIST dataset but with small modifications it can be used for other datasets too. See [Accuracy](#Accuracy) for more info.
 
-# GUI
-![GUI](./data/assets/C-digit-recognition-example-of-usage.jpg)
-
-## Control
-- `left mouse button & drag` to draw
-- `right mouse button` to clear the draw space
-- `middle mouse button or Recognise button in the GUI` to run recognition process
 
-## Processing of the drawn image
-Preprocessing used in [MNIST](http://yann.lecun.com/exdb/mnist/) database: *The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and translating the image so as to position this point at the center of the 28x28 field.*
-
-1. **crop calculation** - crop of the whole draw space is calculated from sides until it reaches non white pixels. After that maximum of the width & height of the cropped image is then taken and `crop_x` & `crop_y` & `crop_w` & `crop_h` is recalculated to preserve image ratio.
-2. **sub image generation** - using previous values and drawn image stored in pixbuf sub image is created
-3. **scaling** - previous cropped image is scaled to the `20x20` image
-4. **conversion into grayscale** - pixbuf is converted into `uint8_t` grayscale image
-5. **adding frame** - `4, 4, 4, 4` frame is added to the `20x20` image resulting into `28x28` image
-6. **computation of the center of mass of the pixels** - is done using mean values accross X & Y
-7. **move of the submatrix** - submatrix (drawn number) is moved in the framed image
-8. **neural network forward propagation** - this preprocessed image is fed to the neural network
 
 # Accuracy
 After `10` epochs of training with batch size `32` the test set accuracy is `97.47 %` which is not bad considering the test error rate in [MNIST database website](http://yann.lecun.com/exdb/mnist/) of the 2-layer NN. Sadly accuracy is not as good in practice as it's in the test data set 🥺.