Lightweight Rebar Segmentation

This code is based on Image Segmentation Keras, but is heavily modified although the instructions to run it should be the same.

This network uses a small number of parameters, is time efficient, and has higher accuracy than other architecture.

Our Architecture:

Our Module:

Comparison in state-of-the-art metrics

Methods	Accuracy (%)	Mean IoU (%)	Params	FPS
ESPnet	99.17	79.39	14,158,566	77
Segnet	99.35	81.40	20,861,480	55
PSPnet	98.57	69.44	25,563,348	47
ENet	99.40	84.30	363,206	71
Our approach	99.49	86.25	10,898,960	91

Image results:

Name	Image
Image
Ground Truth
ESPnet
Segnet
PSPnet
Enet
Our Approach