Implement differentiable binarization and mask erosion

[robertknight]

Implement two changes to make it easier to segment individual words and lines out of the text mask. With the initial U-Net model adjacent words and lines were prone to running together, making them difficult to separate. The plan is to feed the detection results into a model which processes text lines, so the main separation that matters is between adjacent lines.

• Use opencv-python to erode the masks for each word in the target mask. This helps create more clearly defined boundaries between adjacent text instances
• Add differentiable binarization (https://arxiv.org/abs/1911.08947). Not all elements of this paper are implemented, just the core element consisting of the probability / threshold head separation and the differentiable binarization function.

Adding DB introduces parameters and computations into the model which are only needed at training time. To optimize inference speed, an eval_only option has been added to DetectionModel which controls whether training-only parameters are added, and module.training checks have been added in the forwards pass to skip unnecessary computation during inference.

Below are some comparison images of 7a22d8bfc431806736311b6b6a9624cf9d931695 (+erosion, +data augmentation +higher res mask, -db) vs 0f673888c5e6e150dd556199a47e8d4d4b4cec23 (+erosion, +data augmentation +higher res mask, +db). The main difference to note is that a sequence of words are less often combined into the same region in the version with DB.

There is a downside to adding DB, which is that the loss starts out much higher and takes more epochs to come down to a low value (around 0.10), so it might be worth exploring ways to speed up convergence or other ways to increase separation around the boundaries of text elements.

Without differentiable binarization:

With differentiable binarization:

Without differentiable binarization (2):

With differentiable binarization (2) :

[robertknight]

Improving recognition and layout accuracy is currently the main focus, so this will be revisited later.