diff transformations/thick.py @ 11:dbc806d025a2

Added a thick.py script defining a Thick class transforming randomly the thickness of the characters
author Xavier Glorot <glorotxa@iro.umontreal.ca>
date Wed, 27 Jan 2010 19:14:37 -0500
parents
children a25474d4d34f
line wrap: on
line diff
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/transformations/thick.py	Wed Jan 27 19:14:37 2010 -0500
@@ -0,0 +1,176 @@
+#!/usr/bin/python
+# coding: utf-8
+
+'''
+Simple implementation of random thickness deformation using morphological
+operation of scipy.
+Only one morphological operation applied (dilation or erosion), the kernel is random
+out of a list of 11 symmetric kernels.
+
+Author: Xavier Glorot
+
+Usage:
+'''
+
+import scipy.ndimage.morphology
+import numpy as N
+
+
+class Thick():
+    def __init__(self,complexity = 1):
+        #---------- private attributes
+        self.__nx__ = 32
+        self.__ny__ = 32
+        self.__erodemax__ = 4
+        self.__dilatemax__ = 11
+        self.__structuring_elements__ = [N.asarray([[1,1]]),N.asarray([[1],[1]]),\
+                                        N.asarray([[1,1],[1,1]]),N.asarray([[0,1,0],[1,1,1],[0,1,0]]),\
+                                        N.asarray([[1,1,1],[1,1,1]]),N.asarray([[1,1],[1,1],[1,1]]),\
+                                        N.asarray([[1,1,1],[1,1,1],[1,1,1]]),\
+                                        N.asarray([[1,1,1,1],[1,1,1,1],[1,1,1,1]]),\
+                                        N.asarray([[1,1,1],[1,1,1],[1,1,1],[1,1,1]]),\
+                                        N.asarray([[0,0,1,0,0],[0,1,1,1,0],[1,1,1,1,1],[0,1,1,1,0],[0,0,1,0,0]]),\
+                                        N.asarray([[1,1,1,1],[1,1,1,1]]),N.asarray([[1,1],[1,1],[1,1],[1,1]])]
+        #------------------------------------------------
+        
+        #---------- generation parameters
+        self.erodenb = N.ceil(complexity * self.__erodemax__)
+        self.dilatenb = N.ceil(complexity * self.__dilatemax__)
+        self.Perode = self.erodenb / (self.dilatenb + self.erodenb + 1.0)
+        self.Pdilate = self.dilatenb / (self.dilatenb + self.erodenb + 1.0)
+        assert (self.Perode + self.Pdilate <= 1) & (self.Perode + self.Pdilate >= 0)
+        assert (complexity >= 0) & (complexity <= 1)
+        #------------------------------------------------
+    
+    def _get_current_parameters(self):
+        return [self.erodenb, self.dilatenb, self.Perode, self.Pdilate]
+    
+    def get_settings_names(self):
+        return ['erodenb','dilatenb','Perode','Pdilate']
+    
+    def regenerate_parameters(self, complexity):
+        self.erodenb = N.ceil(complexity * self.__erodemax__)
+        self.dilatenb = N.ceil(complexity * self.__dilatemax__)
+        self.Perode = self.erodenb / (self.dilatenb + self.erodenb + 1.0)
+        self.Pdilate = self.dilatenb / (self.dilatenb + self.erodenb + 1.0)
+        assert (self.Perode + self.Pdilate <= 1) & (self.Perode + self.Pdilate >= 0)
+        assert (complexity >= 0) & (complexity <= 1)
+        return self._get_current_parameters()
+    
+    def transform_1_image(self,image,genparam_save = None):
+        P = N.random.uniform()
+        
+        if P>1-(self.Pdilate+self.Perode):
+            maxi = float(N.max(image))
+            mini = float(N.min(image))
+            
+            if maxi>1.0:
+                image=image/maxi
+            
+            if P>1-(self.Pdilate+self.Perode)+self.Perode:
+                nb=N.random.randint(self.dilatenb)
+                trans=scipy.ndimage.morphology.grey_dilation\
+                        (image,size=self.__structuring_elements__[nb].shape,structure=self.__structuring_elements__[nb])
+                meth = 'dilate'
+            else:
+                nb=N.random.randint(self.erodenb)
+                trans=scipy.ndimage.morphology.grey_erosion\
+                        (image,size=self.__structuring_elements__[nb].shape,structure=self.__structuring_elements__[nb])
+                meth = 'erode'
+            
+            #------renormalizing
+            maxit = N.max(trans)
+            minit = N.min(trans)
+            trans= numpy.asarray((trans - (minit+mini)) / (maxit - (minit+mini)) * maxi,dtype=image.dtype)
+            #--------
+            if genparam_save is not None:
+                genparam_save.update({'Thick':{'meth':meth,'nb':nb}})
+            return trans
+        else:
+            meth = 'nothing'
+            nb = 0
+            if genparam_save is not None:
+                genparam_save.update({'Thick':{'meth':meth,'nb':nb}})
+            return image
+    
+    def transform_image(self,image,genparam_save = None):
+        if image.ndim == 2:
+            newimage = N.reshape(image,(image.shape[0],self.__nx__,self.__ny__))
+            for i in range(image.shape[0]):
+                if genparam_save is not None:
+                    newimage[i,:,:] = self.transform_1_image(newimage[i,:,:],genparam_save[i])
+                else:
+                    newimage[i,:,:] = self.transform_1_image(newimage[i,:,:])
+            return N.reshape(newimage,image.shape)
+        else:
+            newimage = N.reshape(image,(self.__nx__,self.__ny__))
+            if genparam_save is not None:
+                newimage = self.transform_1_image(newimage,genparam_save)
+            else:
+                newimage = self.transform_1_image(newimage)
+            return N.reshape(newimage,image.shape)
+
+
+
+
+#test on NIST (you need pylearn and access to NIST to do that)
+
+if __name__ == '__main__':
+    
+    from pylearn.io import filetensor as ft
+    import copy, numpy
+    import pygame
+    import time
+    datapath = '/data/lisa/data/nist/by_class/'
+    f = open(datapath+'digits/digits_train_data.ft')
+    d = ft.read(f)
+    
+    pygame.surfarray.use_arraytype('numpy')
+    
+    pygame.display.init()
+    screen = pygame.display.set_mode((8*2*32,8*32),0,8)
+    anglcolorpalette=[(x,x,x) for x in xrange(0,256)]
+    screen.set_palette(anglcolorpalette)
+    
+    MyThick = Thick()
+    
+    #debut=time.time()
+    #MyThick.transform_image(d)
+    #fin=time.time()
+    #print '------------------------------------------------'
+    #print d.shape[0],' images transformed in :', fin-debut, ' seconds'
+    #print '------------------------------------------------'
+    #print (fin-debut)/d.shape[0]*1000000,' microseconds per image'
+    #print '------------------------------------------------'
+    #print MyThick.get_settings_names()
+    #print MyThick._get_current_parameters()
+    #print MyThick.regenerate_parameters(0)
+    #print MyThick.regenerate_parameters(0.5)
+    #print MyThick.regenerate_parameters(1)
+    for i in range(10000):
+        a=d[i,:]
+        b=N.asarray(N.reshape(a,(32,32))).T
+        
+        new=pygame.surfarray.make_surface(b)
+        new=pygame.transform.scale2x(new)
+        new=pygame.transform.scale2x(new)
+        new=pygame.transform.scale2x(new)
+        new.set_palette(anglcolorpalette)
+        screen.blit(new,(0,0))
+        
+        dd={}
+        c=MyThick.transform_image(a,dd)
+        b=N.asarray(N.reshape(c,(32,32))).T
+        
+        new=pygame.surfarray.make_surface(b)
+        new=pygame.transform.scale2x(new)
+        new=pygame.transform.scale2x(new)
+        new=pygame.transform.scale2x(new)
+        new.set_palette(anglcolorpalette)
+        screen.blit(new,(8*32,0))
+        
+        pygame.display.update()
+        print dd
+        raw_input('Press Enter')
+    
+    pygame.display.quit()
\ No newline at end of file