week3 & 4 : d3bug.mkv # Learning Based Image Colorization
Team Members : Gökay Atay, Bugrahan Akbulut
Previously on our project, we were following the research Color Image Colorization. They have VGGish network to colorize image which is trained on ImageNet Dataset. They implemented their model and related operations on Caffe Framework. So we decided to create a model with ResNet to complete colorization task.
Due to some errors, we could not run their train code so we decided to keep digging. Then we found another research which is called Automatic Image Colorization. In Automatic Image Colorization, they used to relatively small network to Colorful Image Colorization by using Resnet-18 Gray and Deconvolution and UpSampling. They trained their own model on Places365.
Here some demo result on their research:
After finding these research we have to move our idea because it is already implemented and researched, we thoughtthat we have to find a new idea in other words have to debug our idea. After lots of thinking, we came up with an idea : Creating our own colorization model with using DenseNet121.
So What Is DenseNet?
In the ResNet network, forces learn residuals by merge previous layers into next one. In contrast, DenseNet paper proposes concatenating outputs from the previous layers instead of using the summation.
In Dense Convolutional Network (DenseNet) each layer connects to every other layer in feed-forward fashion. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters.
Effect of this dense connectivity pattern is that it requires fewer parameters than traditional convolutional networks, as there is no need to relearn redundant feature maps.
Finally, we thought to merge one or two dense blocks(pretrained) followed by deconvolution and upsampling layers.
Soon we will share how our network is shown…
