Based on image-search.ipynb and image-tsne.ipynb

• Solve the alpha channel issue for feature extraction; directly pull images from the API
• "This is due to the fact that eps does not know about transparency and the default background color for the rasterization was (0, 0, 0, 0) (black which is fully transparent)." --- continuing error in analysis + expanding features

def load_image(path):
     img = Image.open(path)
     img = scipy.misc.imresize(np.array(img), (224, 224), interp='bicubic')
     img[:,:,0] = 255.0-img[:,:,3]
     img[:,:,1] = 255.0-img[:,:,3]
     img[:,:,2] = 255.0-img[:,:,3]
     img = img[:,:,:3]
     x = image.img_to_array(img)
     x = np.expand_dims(x, axis=0)
     x = preprocess_input(x)
     return img, x

<ipython-input-22-6fd16072f6c2> in load_image(path)
      8     img = Image.open(path)
      9     img = scipy.misc.imresize(np.array(img), (224, 224), interp='bicubic')
---> 10     img[:,:,0] = 255.0-img[:,:,3]
     11     img[:,:,1] = 255.0-img[:,:,3]
     12     img[:,:,2] = 255.0-img[:,:,3]

 IndexError: index 3 is out of bounds for axis 2 with size 3 

• Test with more than thousands images (max 50 per a call)

auth = OAuth1("API KEY", "API KEY2")
 endpoint = "https://api.thenounproject.com/icons/{term}?page=2"
 response = requests.get(endpoint, auth=auth)
 with open('./bicycle-p2.json', 'w') as results_file:
     json.dump(response.json(), results_file)

• Matching across different terms
• The dissimilarity metric I used is cosine distance (do the vectors point in the same direction?). Also consider:
  • Euclidean distance (are the vectors similar in direction and magnitude?)
  • Hamming distance (are the vectors roughly similar?)
  • Both from https://github.com/Jack000/fontjoy

Using a pre-trained model in Keras, e.g., VGG, to extract the feature of a given image;

• https://medium.com/@franky07724_57962/using-keras-pre-trained-models-for-feature-extraction-in-image-clustering-a142c6cdf5b1
• https://github.com/ml4a/ml4a-guides/blob/master/notebooks/image-search.ipynb
• transparency in png file?
• calling multiple pages of icon term