Segmentation: Identify the relevant parts of an image
Image in, image out, thus an autoencoder problem
• Very important for medical images
• Emphasize areas where the expert used should look for signs of decease
(or other)
• Improve classi cation: 1) segment, 2) crop, 3) run usual CNN classi cation
fi
Illustrations: Chollet
Clarifying a point in the
segmentation example
Your teacher could not answer, but now he can – see the next slides
Example: Segmenting the animals in an image
chapter09_part01_image-segmentation
?????
Example: Segmenting the animals in an image
chapter09_part01_image-segmentation
?????
Model structure: auto-encoder
Explanation:
• the outputs layer returns an "prob. distr." for each pixel
• a good choice for training the model,
• but it is not a proper segmentation
However, the "prob. distr." is changed into a single value in
the visualization:
def display_mask(pred):
mask = np.argmax(pred, axis=-1)
mask *= 127
plt.axis("off")
plt.imshow(mask)
Results
Results