README.md

Image_Segmentation (VGG+FCN8 and UNet)

1. VGG16 + FCN8 in the folder VFF_FCN8.
   This work uses VGG16 from scratch as a encoder and FCN8 as decoder to build a image segmentation model. The diagram of the network shown as,

The dataset uses dataset1 prepared by divamgupta.

The dataset contains folders for images and annotations. The images contain the video frames while the annotations contain the pixel-wise label maps. Each label map has the shape (height, width , 1) with each point in this space denoting the corresponding pixel's class. Classes are in the range [0, 11] (i.e. 12 classes) and the pixel labels correspond to these classes
data_processing.py contains functions of pre-processing dataset
image_processing.py includes functions of image visulization.

The VGG16 weights can be downloaded from VGG16_weight

Here is the result after fitting the model with 170 epochs receiving 0.85 accuracy(IOU and Dice were calculated to assess the performance):

2. UNet in the folder UNet. The model will use the dataset Oxford Pets - IIT dataset, which is included in TensorFlow Datasets already. The UNet diagram is shown as,

The validation accuracy after 20 epochs reaches 0.89, and one of the resulting examples is shown as,