Semantic segmentation project aims to generate mask images identifying salt deposits through a deep learning model.
---------------------------------------TF_implementation--------------------------------------
Input: 128x128, grey scale image.
Augmentations
Brightness , vflip, hflip
Base model
U-net model, with last block modified by several valid convolutions followed by a 1×1 convolution to squeeze tensors' channels and end with a 101×101×1 image.
ScSE (Spatial-Channel Squeeze & Excitation) both in Encoder and Decoder
dropout was applied in the second stage of training with keep prob value 0.5.
Learning
Batch size =15 (maximum available batch size), Adam optimizer, evaluation metric is mean intersection over union, learning rate assigned manually at every training stage.
---------------------------------------Keras implementation--------------------------------------
Input: 101x101, RGB scale image padded with reflect mode to 128x128.
Base model
U-net model, resnet34 encoder with last block excited with feature pyramid attention network
ScSE (Spatial-Channel Squeeze & Excitation) in Decoder
dropout was applied in the second stage of training with keep prob value 0.5.
Learning
Batch size =12 (maximum available batch size), Adam optimizer, evaluation metric is mean intersection over union, learning rate reduced on plateau.
better results than tensorflow (0.84 vs 0.76)