Letter Recognition

letterrecognition.py

from pybrain.structure import RecurrentNetwork, FeedForwardNetwork, FullConnection, LinearLayer, SigmoidLayer
from pybrain.datasets import SupervisedDataSet
from pybrain.supervised.trainers import BackpropTrainer
from math import log,exp
import cv,cv2
import os
import random
import ImageFont, Image, ImageDraw
import time
import numpy as np
import pickle

samplecount = 10 # Number of randomly sized samples per character per font
characters = 'abcdefghijklmnopqrstuvwxyz'
fontpath = "/usr/share/fonts/"
fontnames = ["liberation-fonts/LiberationMono-Bold.ttf","liberation-fonts/LiberationSerif-BoldItalic.ttf","liberation-fonts/LiberationMono-Italic.ttf","liberation-fonts/LiberationSerif-Bold.ttf","liberation-fonts/LiberationMono-Regular.ttf","liberation-fonts/LiberationSerif-Italic.ttf","liberation-fonts/LiberationSans-Bold.ttf","liberation-fonts/LiberationMono-BoldItalic.ttf","liberation-fonts/LiberationSerif-Regular.ttf","liberation-fonts/LiberationSans-Regular.ttf","liberation-fonts/LiberationSans-BoldItalic.ttf","liberation-fonts/LiberationSans-Italic.ttf","libertine-ttf/LinBiolinum_aSI.ttf","libertine-ttf/LinBiolinum_aRL.ttf","libertine-ttf/LinBiolinum_RI.ttf","libertine-ttf/LinBiolinum_aW.ttf","libertine-ttf/LinLibertine_RZ.ttf","libertine-ttf/LinBiolinum_aUB.ttf","libertine-ttf/LinLibertine_RB.ttf","libertine-ttf/LinLibertine_aRL.ttf","libertine-ttf/LinBiolinum_R.ttf","libertine-ttf/LinLibertine_aBS.ttf","libertine-ttf/LinBiolinum_K.ttf","libertine-ttf/LinLibertine_DR.ttf","libertine-ttf/LinBiolinum_RB.ttf","libertine-ttf/LinLibertine_RBI.ttf","libertine-ttf/LinLibertine_aBL.ttf","libertine-ttf/LinLibertine_R.ttf","libertine-ttf/LinBiolinum_aWI.ttf","libertine-ttf/LinBiolinum_aSB.ttf","libertine-ttf/LinBiolinum_aBL.ttf","libertine-ttf/LinBiolinum_aUI.ttf","libertine-ttf/LinLibertine_RI.ttf","libertine-ttf/LinLibertine_aSZI.ttf","libertine-ttf/LinLibertine_aDRL.ttf","libertine-ttf/LinLibertine_I.ttf","libertine-ttf/LinLibertine_aZL.ttf","libertine-ttf/LinBiolinum_aU.ttf","libertine-ttf/LinLibertine_RZI.ttf","libertine-ttf/LinBiolinum_aWB.ttf","libertine-ttf/LinLibertine_aSI.ttf","libertine-ttf/LinLibertine_aS.ttf","libertine-ttf/LinBiolinum_aS.ttf","libertine-ttf/LinLibertine_aDRS.ttf","dejavu/DejaVuSerifCondensed-Italic.ttf","dejavu/DejaVuSerif.ttf","dejavu/DejaVuSansMono.ttf","dejavu/DejaVuSans.ttf","dejavu/DejaVuSansCondensed.ttf","dejavu/DejaVuSansCondensed-Bold.ttf","dejavu/DejaVuSerif-Bold.ttf","dejavu/DejaVuSansCondensed-BoldOblique.ttf","dejavu/DejaVuSans-ExtraLight.ttf","dejavu/DejaVuSans-BoldOblique.ttf","dejavu/DejaVuSansMono-Oblique.ttf","dejavu/DejaVuSerif-Italic.ttf","dejavu/DejaVuSansMono-Bold.ttf","dejavu/DejaVuSans-Bold.ttf","dejavu/DejaVuSerifCondensed-BoldItalic.ttf","dejavu/DejaVuSerif-BoldItalic.ttf","dejavu/DejaVuSansMono-BoldOblique.ttf","dejavu/DejaVuSerifCondensed.ttf","dejavu/DejaVuSerifCondensed-Bold.ttf","dejavu/DejaVuSans-Oblique.ttf","dejavu/DejaVuSansCondensed-Oblique.ttf","corefonts/verdanaz.ttf","corefonts/timesbi.ttf","corefonts/couri.ttf","corefonts/impact.ttf","corefonts/arial.ttf","corefonts/georgiai.ttf","corefonts/comic.ttf","corefonts/ariblk.ttf","corefonts/courbi.ttf","corefonts/georgia.ttf","corefonts/times.ttf","corefonts/verdanai.ttf","corefonts/courbd.ttf","corefonts/trebucit.ttf","corefonts/trebucbi.ttf","corefonts/comicbd.ttf","corefonts/webdings.ttf","corefonts/timesbd.ttf","corefonts/georgiab.ttf","corefonts/cour.ttf","corefonts/timesi.ttf","corefonts/verdana.ttf","corefonts/arialbi.ttf","corefonts/verdanab.ttf","corefonts/trebuc.ttf","corefonts/andalemo.ttf","corefonts/ariali.ttf","corefonts/trebucbd.ttf","corefonts/georgiaz.ttf","corefonts/arialbd.ttf"]
fonts = [os.path.join(fontpath,f) for f in fontnames]

network = FeedForwardNetwork(name="Letter_Recognition")

inputLayer = LinearLayer(400, name="Input")
hiddenLayer = SigmoidLayer(7, name="Pre-Classification")
hiddenLayer2 = SigmoidLayer(7, name="Post-Classification")
outputLayer = LinearLayer(1, name="Output")

network.addInputModule(inputLayer)
network.addModule(hiddenLayer)
network.addModule(hiddenLayer2)
network.addOutputModule(outputLayer)

c1 = FullConnection(inputLayer, hiddenLayer, name="Input_to_Hidden")
c5 = FullConnection(hiddenLayer, hiddenLayer2, name="Hidden_to_Hidden2")
c2 = FullConnection(hiddenLayer2, outputLayer, name="Hidden2_to_Output")
#c3 = FullConnection(hiddenLayer, hiddenLayer, name="Recurrent_Connection")
#c4 = FullConnection(hiddenLayer2, hiddenLayer2, name="Recurrent_Connection2")
 
network.addConnection(c1)
network.addConnection(c5)
#network.addRecurrentConnection(c3)
network.addConnection(c2)

network.sortModules()

print network

try:
    ds = pickle.load(open('_letterds'))
except:
    ds = SupervisedDataSet(400, 1)
    for font_path in fonts:
        print(font_path)
        WIDTH = 20
        HEIGHT = 20
        for char in characters:
            for _ in range(samplecount):
                pt = random.randint(8,14)
                try:
                    font = ImageFont.truetype(font_path,pt,encoding='unic')
                except:
                    break
                size_x, size_y = font.getsize(char)
                x = (WIDTH-size_x)/2
                y = (HEIGHT-size_y)/2
                im = Image.new('RGB',size=(HEIGHT,WIDTH))
                draw = ImageDraw.Draw(im)
                draw.text((x,y),char,font=font)
                cv_img = np.array(im)  # RGB image
                cv_img = cv_img[:, :, ::-1].copy()
                char2 = cv2.Canny(cv_img, 100, 200)
                pixels = []
                for row in char2:
                    for pixel in row:
                        if pixel == 255:
                            pixels.append(1)
                        else:
                            pixels.append(0)
                ds.addSample(pixels, [(characters.index(char)+1)/26.])
                #cv2.namedWindow("CHAR")
                #cv2.imshow("CHAR",cv_img)
                #cv2.waitKey(0)
                #cv2.destroyAllWindows()
                #time.sleep(5000)
    pickle.dump(ds, open('_letterds','w'))
    
print ds

trainer = BackpropTrainer(network, ds, momentum=0.99, learningrate=0.01)

print "\nInitial weights: ", network.params
 
max_error = 1e-7
error, count = 1, 1000
#Train
while abs(error) >= max_error and count > 0:
    error = trainer.train()
    count = count - 1
    print error

print network.params
pickle.dump(network, open('_network','w'))