基於深度強化學習神經網路之自動化裂縫分割與偵測

Abstract

傳統的裂縫檢測依賴人工，耗時且需要大量勞力，近年來由於AI技術的迅速發展，有許多自動化裂縫偵測的研究出現，其中裂縫分割為近期發展的重點，透過卷積神經網路，能將裂縫的特徵萃取出來，把裂縫與影像中的背景進行分類，重構成分割後的影像，此方法是像素等級的裂縫辨識，然而，裂縫分割缺乏了圖像蒐集的自動化，迄今為止的裂縫分割研究多是透過現場拍照或網路上的圖片做為資料集，代表即使有了這項技術，依然要透過人力蒐集照片。 因此，本研究提出了基於深度雙Q類神經網路 (Double Deep Q-Network, DDQN) 的裂縫偵測模型，它是一種強化學習演算法，透過與環境的互動建構出一套行為模式，使機器人能自動抓取裂縫影像，本研究以預訓練的U-Net影像分割模型即時的偵測裂縫並提供獎勵，DDQN透過這項獎勵系統更新網路並不斷優化自己的行為，藉此偵測出最多的裂縫，本研究以過往裂縫分割研究使用的影像資料集作為模擬環境，結果顯示，在訓練環境中，機器人展現了它對行為選擇的準確性，在偵測到一部分的裂縫後便能順著裂縫的輪廓偵測出整條裂縫，且在測試環境，機器人也有良好的適應性，代表此強化學習架構的成功。

Traditional crack detection relies on manual inspection, which is time consuming and labor intensive. In recent years, due to the rapid development of artificial intelligence (AI), there has been many researches on the automation of crack detection. Among these studies, crack segmentation is the most advanced topic. Through convolutional neural networks, features of cracks are extracted, allowing the segmentation of cracks against its background. This method represents pixel-level classification. However, crack segmentation lacks automation in data collection. Most existing crack segmentation studies rely on photography or online images as datasets, indicating that manual efforts are still needed to gather photos. Therefore, this study proposes a crack detection model based on the deep double Q neural network (DDQN). It is a reinforcement learning algorithm for the agent to learn through interactions with the environment, enabling the robot to automatically capture crack images. The study utilizes a pre-trained U-Net image segmentation model for real-time crack detection, providing rewards. The DDQN continuously optimizes its behavior through the reward system, thereby trained to capture the maximum number of cracks. Using the image dataset employed in past crack segmentation studies as a simulated environment, the results demonstrate the robot''s accuracy in behavior selection in the training environment. After detecting a portion of the crack, it can seamlessly gather images along the entire crack. In the testing environment, the robot maintains this capability, indicating the success of the reinforcement learning framework in automating the collection of crack images.