由於台灣氣候相當潮溼以及炎熱,使得鋼橋等等的鋼結構設施相當容易生鏽,為此,希望能夠建立一個有效率且客觀的鏽蝕偵測工具,來降低鏽蝕例行檢測的人力、時間成本,以及提高檢測成果的可信度。 過往的鏽蝕偵測相關研究主要著重於利用邊緣演算法對鏽蝕的邊緣進行刻劃,進而計算鏽蝕面積,主要會面臨到的問題在於能夠辨識的鏽蝕種類受到限制,以及難以辨識異物與鏽蝕之間的差距。 本研究引入目前流行的深度學習技術,利用全卷積神經網絡建立辨識模型對鏽蝕圖像進行語義分割,來改善以往辨識方式中面臨到的限制。
The weather in Taiwan is hot and wet, which makes steel structures, such as steel bridges, easy to rust. Therefore, establishing an effective method for corrosion detection is important in maintaining infrastructure’s “health” and reducing the corresponding lifecycle cost. Browsing past research efforts, there were a number of image processing techniques (IPTs) proposed for quick and effective recognition. A crucial issue has been on distinguishing real corrosion from noises or patterns which look like rust. Also, there is a limit to the type of the rust which can be detected. In this research, the fully convolutional neural network named U-Net will be explored to establish an image semantic segmentation model, which will be able to deal with a wide range of rust type.