Deep Learning in Semantic Segmentation of Rust in Images

Abstract

Rust detection is an essential topic in many areas, especially in telecommunication, which needs effective systems to segment and recognize rust on power electric towers, antenna. Our exclusive architecture use is based on a fully convolutional neural network for semantic segmentation and composed of Densenet encoder PSP intermediate layers and two skip connections upsample layers. The code written in Python used Pytorch libraries to compute and categorize the images. Comparing between models such as E-Net, U-Net, FCN, we have received our highest FCN (Fully Convolutional Neural) model for the most stable ratio of IoU (Intersection over Union) in 3 models stated with mean scores are 58.1 for origin images and 61.8 for background removal. With the results, we will contribute to detect rust on electric poles in time to avoid rust-causing serious consequences.