Working with pretty damn good accuracy

mini_proj/Load_Images.py

@@ -7,6 +7,8 @@ from matplotlib import pyplot as plt
 import math
 import cv2
 from skimage import color, exposure, transform
+import skimage as sk
+import random
 
 '''
 Defines some basic preprocessing for the images
@@ -17,8 +19,29 @@ def preprocess(img):
     hsv[:, :, 2] = exposure.equalize_hist(hsv[:, :, 2])
     img = color.hsv2rgb(hsv)
 
-    img = img/255                           # Scaling images down to values of 0-255
-    img = np.rollaxis(img, -1)              # rolls colour axis to 0 
+    #img = img/255                           # Scaling images down to values of 0-255
+
+    return img
+
+'''
+Augments an image (horizontal reflection, hue, contrast, saturation adjustment) to expand on the dataset
+Code was adapted from Medium.com/@thimblot (Thomas Himblot)
+'''
+def augment(img):
+    # Randomly rotate the image
+    random_degree = random.uniform(-25, 25)         # Random degree of rotation between 25% on the left and 25% on the right
+    img = sk.transform.rotate(img, random_degree)
+
+    # Add random noise to the image
+    img = sk.util.random_noise(img)
+
+    # Randomly (25% chance) flips the image horizontally (by inverting the image dimensions)
+    if random.randint(0, 4) == 1:
+        img = img[:, ::-1]
+
+    # Randomly (25% chance) inverts the intensity range of the image
+    if random.randint(0, 4) == 1:
+        img = sk.util.invert(img)
 
     return img
 
@@ -30,27 +53,40 @@ def gen_data(w_path, n_w_path):
     imgs_raw = []           # Images
     imgs_lbl = []           # Image labels
 
-    i = 0
+    w = 0
     for image_name in waldo_file_list:
         pic = cv2.imread(os.path.join(w_path, image_name))              # NOTE: cv2.imread() returns a numpy array in BGR not RGB
         pic = preprocess(pic)
-        
-        imgs_raw.append(pic)
+
+        pic_roll = np.rollaxis(pic, -1)                                 # rolls colour axis to 0 
+        imgs_raw.append(pic_roll)
         imgs_lbl.append(1)                                              # Value of 1 as Waldo is present in the image
 
-        print('Completed: {0}/{1} Waldo images'.format(i+1, total_w))
-        i += 1
+        for j in range(0, 10):
+            pic = augment(pic)
+            pic_roll = np.rollaxis(pic, -1)                                 # rolls colour axis to 0 
+            imgs_raw.append(pic_roll)
+            imgs_lbl.append(1) 
 
-    i = 0
+
+        print('Completed: {0}/{1} Waldo images'.format(w+1, total_w))
+        w += 1
+
+    nw = 0
     for image_name in not_waldo_file_list:
         pic = cv2.imread(os.path.join(n_w_path, image_name))
         pic = preprocess(pic)
         
+        pic = np.rollaxis(pic, -1)                                      # rolls colour axis to 0 
         imgs_raw.append(pic)
         imgs_lbl.append(0)
         
-        print('Completed: {0}/{1} non-Waldo images'.format(i+1, total_nw))
-        i += 1
+        print('Completed: {0}/{1} non-Waldo images'.format(nw+1, total_nw))
+        if nw > 50*w:
+            print("Non_Waldo files restricted")
+            break
+        else:
+            nw += 1
 
     ## Randomise and split data into training and test sets
     # Code was modified from code written by: Kyle O'Brien (medium.com/@kylepob61392)

mini_proj/test_nn.py

@@ -8,8 +8,12 @@ test_y = np.load("Waldo_test_lbl.npy")
 test_y = to_categorical(test_y)
 
 f = open("test_output.txt", 'w')
-
+z = 0
 for i in range(0, len(test_y)):
     print(pred_y[i], test_y[i], file=f)
+    # Calculates correct predictions 
+    if pred_y[i][0] == test_y[i][0]:
+        z+=1
 
+print("Accuracy: {}%".format(z/len(test_y)*100))
 f.close()

mini_proj/waldo_model.py

@@ -117,7 +117,7 @@ model = FCN()
 # ensemble_iclf = ImageClassifier(ensemble.RandomForestClassifier)
 
 ## Define training parameters
-epochs = 20             # an epoch is one forward pass and back propogation of all training data
+epochs = 10             # an epoch is one forward pass and back propogation of all training data
 batch_size = 150        # batch size - number of training example used in one forward/backward pass
 # (higher batch size uses more memory, smaller batch size takes more time)
 #lrate = 0.01           # Learning rate of the model - controls magnitude of weight changes in training the NN