Our network takes as input a 3 channel black-and-white image(each channel is repeated) and it gives as output a 5-channel image containing on each channel the mask for each of the website primitive. To finish building the dataset, follow these steps AFTER running the crawler from the corpus directory:
cd corpuspython3 make_corpus.py --scraped-websites-dir /path/to/scraped/websites --class-dirs list-with-scraped-primitive-namesAfter this, the dataset is ready.
In our experiments, we have tried models based on the encoder-decoder architecture, that contain a downsampling and an upsampling path. Currently, in the UI code, we are using frednetv2, but you can go ahead and change it to see the results we have achieved with the others.
In the table below, we have explained the differences between the implementations of UNet, PSPNet and Frednetv2, since the other frednet iterations are similar.
| Name | Details |
|---|---|
| UNet | Based on the implementation of U-Net: Convolutional Networks for Biomedical Image Segmentation. Contains an encoder-decoder path with residual connections. It was initially used to segment images of brain scans, however we adapted it for our purpose. |
| PSPNet | Based on Pyramid Scene Parsing Network. It is similar to the other encoder-decoder architectures, but the residual connections are achieved by concantenating layers of different sizes from the encoder and passing them through the decoder. |
| FrednetV2 | This is an original architecture. We also base our model on the encoder-decoder networks, but we empirically determined the depth needed for our dataset. Another improvement came from using Deconvolution and Checkerboard Artifacts to remove the checkerboard artifacts obtained from the learnable upsampling layers. |