Posit AI Blog: Image segmentation with U-Net

Sure, it’s good when I’ve an image of some object, and a neural community can inform me what sort of object that’s. More realistically, there is perhaps a number of salient objects in that image, and it tells me what they’re, and the place they’re. The latter activity (referred to as object detection) appears particularly prototypical of up to date AI functions that on the identical time are intellectually fascinating and ethically questionable. It’s completely different with the topic of this submit: Successful picture segmentation has loads of undeniably helpful functions. For instance, it’s a sine qua non in drugs, neuroscience, biology and different life sciences.

So what, technically, is picture segmentation, and the way can we practice a neural community to do it?

Image segmentation in a nutshell

Say we now have a picture with a bunch of cats in it. In classification, the query is “what’s that?” and the reply we wish to hear is: “cat.” In object detection, we once more ask “what’s that,” however now that “what” is implicitly plural, and we count on a solution like “there’s a cat, a cat, and a cat, and they’re here, here, and here” (think about the community pointing, via drawing bounding packing containers, i.e., rectangles across the detected objects). In segmentation, we would like extra: We need the entire picture lined by “boxes” – which aren’t packing containers anymore, however unions of pixel-size “boxlets” – or put in another way: We need the community to label each single pixel within the picture.

Here’s an instance from the paper we’re going to speak about in a second. On the left is the enter picture (HeLa cells), subsequent up is the bottom reality, and third is the discovered segmentation masks.

Figure 1: Example segmentation from Ronneberger et al. 2015.

Technically, a distinction is made between class segmentation and occasion segmentation. In class segmentation, referring to the “bunch of cats” instance, there are two attainable labels: Every pixel is both “cat” or “not cat.” Instance segmentation is harder: Here each cat will get their very own label. (As an apart, why ought to that be harder? Presupposing human-like cognition, it wouldn’t be – if I’ve the idea of a cat, as an alternative of simply “cattiness,” I “see” there are two cats, not one. But relying on what a particular neural community depends on most – texture, colour, remoted elements – these duties could differ rather a lot in issue.)

The community structure used on this submit is enough for class segmentation duties and must be relevant to an enormous variety of sensible, scientific in addition to non-scientific functions. Speaking of community structure, how ought to it look?

Introducing U-Net

Given their success in picture classification, can’t we simply use a basic structure like Inception V[n], ResNet, ResNext … , no matter? The drawback is, our activity at hand – labeling each pixel – doesn’t match so effectively with the basic thought of a CNN. With convnets, the concept is to use successive layers of convolution and pooling to construct up characteristic maps of reducing granularity, to lastly arrive at an summary degree the place we simply say: “yep, a cat.” The counterpart being, we lose element data: To the ultimate classification, it doesn’t matter whether or not the 5 pixels within the top-left space are black or white.

In apply, the basic architectures use (max) pooling or convolutions with stride > 1 to attain these successive abstractions – essentially leading to decreased spatial decision.
So how can we use a convnet and nonetheless protect element data? In their 2015 paper U-Net: Convolutional Networks for Biomedical Image Segmentation (Ronneberger, Fischer, and Brox 2015), Olaf Ronneberger et al. got here up with what 4 years later, in 2019, remains to be the preferred strategy. (Which is to say one thing, 4 years being a very long time, in deep studying.)

The thought is stunningly easy. While successive encoding (convolution / max pooling) steps, as normal, cut back decision, the following decoding – we now have to reach at an output of dimension identical because the enter, as we wish to label each pixel! – doesn’t merely upsample from essentially the most compressed layer. Instead, throughout upsampling, at each step we feed in data from the corresponding, in decision, layer within the downsizing chain.

For U-Net, actually an image says greater than many phrases:

Figure 2: U-Net structure from Ronneberger et al. 2015.

At every upsampling stage we concatenate the output from the earlier layer with that from its counterpart within the compression stage. The remaining output is a masks of dimension the unique picture, obtained by way of 1×1-convolution; no remaining dense layer is required, as an alternative the output layer is only a convolutional layer with a single filter.

Now let’s really practice a U-Net. We’re going to make use of the unet bundle that permits you to create a well-performing mannequin in a single line:

remotes::install_github("r-tensorflow/unet")
library(unet)

# takes further parameters, together with variety of downsizing blocks, 
# variety of filters to begin with, and variety of lessons to determine
# see ?unet for more information
mannequin <- unet(input_shape = c(128, 128, 3))

So we now have a mannequin, and it appears like we’ll be eager to feed it 128×128 RGB photographs. Now how can we get these photographs?

The knowledge

To illustrate how functions come up even outdoors the world of medical analysis, we’ll use for instance the Kaggle Carvana Image Masking Challenge. The activity is to create a segmentation masks separating automobiles from background. For our present function, we solely want practice.zip and train_mask.zip from the archive offered for obtain. In the next, we assume these have been extracted to a subdirectory known as data-raw.

Let’s first check out some photographs and their related segmentation masks.

The photographs are RGB-space JPEGs, whereas the masks are black-and-white GIFs.

We break up the info right into a coaching and a validation set. We’ll use the latter to watch generalization efficiency throughout coaching.

knowledge <- tibble(
  img = list.files(right here::right here("data-raw/practice"), full.names = TRUE),
  masks = list.files(right here::right here("data-raw/train_masks"), full.names = TRUE)
)

knowledge <- initial_split(knowledge, prop = 0.8)

To feed the info to the community, we’ll use tfdatasets. All preprocessing will find yourself in a easy pipeline, however we’ll first go over the required actions step-by-step.

Preprocessing pipeline

The first step is to learn within the photographs, making use of the suitable features in tf$picture.

training_dataset <- coaching(knowledge) %>%  
  tensor_slices_dataset() %>% 
  dataset_map(~.x %>% list_modify(
    # decode_jpeg yields a 3d tensor of form (1280, 1918, 3)
    img = tf$picture$decode_jpeg(tf$io$read_file(.x$img)),
    # decode_gif yields a 4d tensor of form (1, 1280, 1918, 3),
    # so we take away the unneeded batch dimension and all however one 
    # of the three (an identical) channels
    masks = tf$picture$decode_gif(tf$io$read_file(.x$masks))[1,,,][,,1,drop=FALSE]
  ))

While developing a preprocessing pipeline, it’s very helpful to verify intermediate outcomes.
It’s straightforward to do utilizing reticulate::as_iterator on the dataset:

$img
tf.Tensor(
[[[243 244 239]
  [243 244 239]
  [243 244 239]
  ...
 ...
  ...
  [175 179 178]
  [175 179 178]
  [175 179 178]]], form=(1280, 1918, 3), dtype=uint8)

$masks
tf.Tensor(
[[[0]
  [0]
  [0]
  ...
 ...
  ...
  [0]
  [0]
  [0]]], form=(1280, 1918, 1), dtype=uint8)

While the uint8 datatype makes RGB values straightforward to learn for people, the community goes to count on floating level numbers. The following code converts its enter and moreover, scales values to the interval [0,1):

training_dataset <- training_dataset %>% 
  dataset_map(~.x %>% list_modify(
    img = tf$image$convert_image_dtype(.x$img, dtype = tf$float32),
    mask = tf$image$convert_image_dtype(.x$mask, dtype = tf$float32)
  ))

To reduce computational cost, we resize the images to size 128x128. This will change the aspect ratio and thus, distort the images, but is not a problem with the given dataset.

training_dataset <- training_dataset %>% 
  dataset_map(~.x %>% list_modify(
    img = tf$image$resize(.x$img, size = shape(128, 128)),
    mask = tf$image$resize(.x$mask, size = shape(128, 128))
  ))

Now, it’s well known that in deep learning, data augmentation is paramount. For segmentation, there’s one thing to consider, which is whether a transformation needs to be applied to the mask as well – this would be the case for e.g. rotations, or flipping. Here, results will be good enough applying just transformations that preserve positions:

random_bsh <- function(img) {
  img %>% 
    tf$image$random_brightness(max_delta = 0.3) %>% 
    tf$image$random_contrast(lower = 0.5, upper = 0.7) %>% 
    tf$image$random_saturation(lower = 0.5, upper = 0.7) %>% 
    # make sure we still are between 0 and 1
    tf$clip_by_value(0, 1) 
}

training_dataset <- training_dataset %>% 
  dataset_map(~.x %>% list_modify(
    img = random_bsh(.x$img)
  ))

Again, we can use as_iterator to see what these transformations do to our images:

Here’s the complete preprocessing pipeline.

create_dataset <- function(data, train, batch_size = 32L) {
  
  dataset <- data %>% 
    tensor_slices_dataset() %>% 
    dataset_map(~.x %>% list_modify(
      img = tf$image$decode_jpeg(tf$io$read_file(.x$img)),
      mask = tf$image$decode_gif(tf$io$read_file(.x$mask))[1,,,][,,1,drop=FALSE]
    )) %>% 
    dataset_map(~.x %>% list_modify(
      img = tf$picture$convert_image_dtype(.x$img, dtype = tf$float32),
      masks = tf$picture$convert_image_dtype(.x$masks, dtype = tf$float32)
    )) %>% 
    dataset_map(~.x %>% list_modify(
      img = tf$picture$resize(.x$img, dimension = form(128, 128)),
      masks = tf$picture$resize(.x$masks, dimension = form(128, 128))
    ))
  
  # knowledge augmentation carried out on coaching set solely
  if (practice) {
    dataset <- dataset %>% 
      dataset_map(~.x %>% list_modify(
        img = random_bsh(.x$img)
      )) 
  }
  
  # shuffling on coaching set solely
  if (practice) {
    dataset <- dataset %>% 
      dataset_shuffle(buffer_size = batch_size*128)
  }
  
  # practice in batches; batch dimension may must be tailored relying on
  # obtainable reminiscence
  dataset <- dataset %>% 
    dataset_batch(batch_size)
  
  dataset %>% 
    # output must be unnamed
    dataset_map(unname) 
}

Training and check set creation now could be only a matter of two perform calls.

training_dataset <- create_dataset(coaching(knowledge), practice = TRUE)
validation_dataset <- create_dataset(testing(knowledge), practice = FALSE)

And we’re prepared to coach the mannequin.

Training the mannequin

We already confirmed the right way to create the mannequin, however let’s repeat it right here, and verify mannequin structure:

mannequin <- unet(input_shape = c(128, 128, 3))
abstract(mannequin)

Model: "mannequin"
______________________________________________________________________________________________
Layer (sort)                   Output Shape        Param #    Connected to                    
==============================================================================================
input_1 (InputLayer)           [(None, 128, 128, 3 0                                          
______________________________________________________________________________________________
conv2d (Conv2D)                (None, 128, 128, 64 1792       input_1[0][0]                   
______________________________________________________________________________________________
conv2d_1 (Conv2D)              (None, 128, 128, 64 36928      conv2d[0][0]                    
______________________________________________________________________________________________
max_pooling2d (MaxPooling2D)   (None, 64, 64, 64)  0          conv2d_1[0][0]                  
______________________________________________________________________________________________
conv2d_2 (Conv2D)              (None, 64, 64, 128) 73856      max_pooling2d[0][0]             
______________________________________________________________________________________________
conv2d_3 (Conv2D)              (None, 64, 64, 128) 147584     conv2d_2[0][0]                  
______________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 32, 32, 128) 0          conv2d_3[0][0]                  
______________________________________________________________________________________________
conv2d_4 (Conv2D)              (None, 32, 32, 256) 295168     max_pooling2d_1[0][0]           
______________________________________________________________________________________________
conv2d_5 (Conv2D)              (None, 32, 32, 256) 590080     conv2d_4[0][0]                  
______________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 16, 16, 256) 0          conv2d_5[0][0]                  
______________________________________________________________________________________________
conv2d_6 (Conv2D)              (None, 16, 16, 512) 1180160    max_pooling2d_2[0][0]           
______________________________________________________________________________________________
conv2d_7 (Conv2D)              (None, 16, 16, 512) 2359808    conv2d_6[0][0]                  
______________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 8, 8, 512)   0          conv2d_7[0][0]                  
______________________________________________________________________________________________
dropout (Dropout)              (None, 8, 8, 512)   0          max_pooling2d_3[0][0]           
______________________________________________________________________________________________
conv2d_8 (Conv2D)              (None, 8, 8, 1024)  4719616    dropout[0][0]                   
______________________________________________________________________________________________
conv2d_9 (Conv2D)              (None, 8, 8, 1024)  9438208    conv2d_8[0][0]                  
______________________________________________________________________________________________
conv2d_transpose (Conv2DTransp (None, 16, 16, 512) 2097664    conv2d_9[0][0]                  
______________________________________________________________________________________________
concatenate (Concatenate)      (None, 16, 16, 1024 0          conv2d_7[0][0]                  
                                                              conv2d_transpose[0][0]          
______________________________________________________________________________________________
conv2d_10 (Conv2D)             (None, 16, 16, 512) 4719104    concatenate[0][0]               
______________________________________________________________________________________________
conv2d_11 (Conv2D)             (None, 16, 16, 512) 2359808    conv2d_10[0][0]                 
______________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTran (None, 32, 32, 256) 524544     conv2d_11[0][0]                 
______________________________________________________________________________________________
concatenate_1 (Concatenate)    (None, 32, 32, 512) 0          conv2d_5[0][0]                  
                                                              conv2d_transpose_1[0][0]        
______________________________________________________________________________________________
conv2d_12 (Conv2D)             (None, 32, 32, 256) 1179904    concatenate_1[0][0]             
______________________________________________________________________________________________
conv2d_13 (Conv2D)             (None, 32, 32, 256) 590080     conv2d_12[0][0]                 
______________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTran (None, 64, 64, 128) 131200     conv2d_13[0][0]                 
______________________________________________________________________________________________
concatenate_2 (Concatenate)    (None, 64, 64, 256) 0          conv2d_3[0][0]                  
                                                              conv2d_transpose_2[0][0]        
______________________________________________________________________________________________
conv2d_14 (Conv2D)             (None, 64, 64, 128) 295040     concatenate_2[0][0]             
______________________________________________________________________________________________
conv2d_15 (Conv2D)             (None, 64, 64, 128) 147584     conv2d_14[0][0]                 
______________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTran (None, 128, 128, 64 32832      conv2d_15[0][0]                 
______________________________________________________________________________________________
concatenate_3 (Concatenate)    (None, 128, 128, 12 0          conv2d_1[0][0]                  
                                                              conv2d_transpose_3[0][0]        
______________________________________________________________________________________________
conv2d_16 (Conv2D)             (None, 128, 128, 64 73792      concatenate_3[0][0]             
______________________________________________________________________________________________
conv2d_17 (Conv2D)             (None, 128, 128, 64 36928      conv2d_16[0][0]                 
______________________________________________________________________________________________
conv2d_18 (Conv2D)             (None, 128, 128, 1) 65         conv2d_17[0][0]                 
==============================================================================================
Total params: 31,031,745
Trainable params: 31,031,745
Non-trainable params: 0
______________________________________________________________________________________________

The “output shape” column reveals the anticipated U-shape numerically: Width and top first go down, till we attain a minimal decision of 8x8; they then go up once more, till we’ve reached the unique decision. At the identical time, the variety of filters first goes up, then goes down once more, till within the output layer we now have a single filter. You also can see the concatenate layers appending data that comes from “below” to data that comes “laterally.”

What must be the loss perform right here? We’re labeling every pixel, so every pixel contributes to the loss. We have a binary drawback – every pixel could also be “car” or “background” – so we would like every output to be near both 0 or 1. This makes binary_crossentropy the enough loss perform.

During coaching, we hold observe of classification accuracy in addition to the cube coefficient, the analysis metric used within the competitors. The cube coefficient is a method to measure the proportion of right classifications:

cube <- custom_metric("cube", perform(y_true, y_pred, clean = 1.0) {
  y_true_f <- k_flatten(y_true)
  y_pred_f <- k_flatten(y_pred)
  intersection <- k_sum(y_true_f * y_pred_f)
  (2 * intersection + clean) / (k_sum(y_true_f) + k_sum(y_pred_f) + clean)
})

mannequin %>% compile(
  optimizer = optimizer_rmsprop(lr = 1e-5),
  loss = "binary_crossentropy",
  metrics = checklist(cube, metric_binary_accuracy)
)

Fitting the mannequin takes a while – how a lot, in fact, will rely in your {hardware}. But the wait pays off: After 5 epochs, we noticed a cube coefficient of ~ 0.87 on the validation set, and an accuracy of ~ 0.95.

Predictions

Of course, what we’re finally considering are predictions. Let’s see just a few masks generated for objects from the validation set:

batch <- validation_dataset %>% as_iterator() %>% iter_next()
predictions <- predict(mannequin, batch)

photographs <- tibble(
  picture = batch[[1]] %>% array_branch(1),
  predicted_mask = predictions[,,,1] %>% array_branch(1),
  masks = batch[[2]][,,,1]  %>% array_branch(1)
) %>% 
  sample_n(2) %>% 
  map_depth(2, perform(x) {
    as.raster(x) %>% magick::image_read()
  }) %>% 
  map(~do.name(c, .x))


out <- magick::image_append(c(
  magick::image_append(photographs$masks, stack = TRUE),
  magick::image_append(photographs$picture, stack = TRUE), 
  magick::image_append(photographs$predicted_mask, stack = TRUE)
  )
)

plot(out)

Figure 3: From left to proper: floor reality, enter picture, and predicted masks from U-Net.

Conclusion

If there have been a contest for the very best sum of usefulness and architectural transparency, U-Net will surely be a contender. Without a lot tuning, it’s attainable to acquire respectable outcomes. If you’re capable of put this mannequin to make use of in your work, or when you have issues utilizing it, tell us! Thanks for studying!