以實例分割與主動學習探討鋼筋間距辨識

Rebar Spacing Recognition with Instance Segmentation and Active Learning

莊仕杰(Shih-Jie Chuang)

指導教授 ： 陳俊杉

國立臺灣大學/工學院/土木工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103220

摘要

本論文透過實例分割建立一創新的鋼筋間距查驗方法並以六項主動學習演算法探討深度學習模型在兩工地場域的標註效率。本研究所提出的量測方法偵測在影像中所有可能的鋼筋間距特徵並藉由索引三維座標計算歐幾里得距離，故得以接收更全面性的量測結果。由於標註成本十分昂貴，模型基於樣本的不確定性選擇最具有價值的資料來減輕標註負擔，在有限的預算內提供一資料標註與模型訓練的流程於技術導入之使用為本研究的主要目標。
本研究首先透過在Steelscapes資料集訓練實例分割模型以偵測影像中的關鍵特徵，包含兩個重點類別：交點與間距。在初始的預鑄樑測試集中，可達到31.006%的平均精度（AP）、0.166公分的絕對誤差與1.388%的相對誤差。除此之外，為了分析模型於不同情境下的測試表現，依照人工分群的屬性以兩個實驗環境作為目標資料集。訓練結果指出改變鋼筋組立的測試情境會大幅降低預測表現，因此透過六種基於物件偵測的查詢策略改善深度學習模型於目標測試集的表現，分別為三種不同聚合方式的熵估計與基於Monte Carlo (MC) dropout的三種分歧量測方法。以平均相對熵估計分歧的RoI matching方法在兩實驗中皆具有最佳的效率表現，分別在固定標註預算內達到81.790%與92.269%的實驗基準表現。最後，主動學習演算法結合人工分群的在連續訓練策略中提供一標準的標註與訓練流程使新資料的效用最大化。

關鍵字

鋼筋間距查驗 ； 實例分割 ； 主動學習 ； 深度學習 ； 彩色深度影像

並列摘要

In this thesis, we present a novel approach for rebar spacing inspection using instance segmentation, and six active learning algorithms to explore the labeling efficiency of the deep learning model in two construction sites. The proposed measurement method detects all possible rebar spacing features in RGB-D images and calculates the Euclidean distance based on the indexed 3D coordinates, thus boosting the spacing inspection progress. Owing to the super-expensive cost of annotation, the model selects the most valuable instances based on their uncertainty to alleviate the burden of labeling. The primary objective of this research is to provide a process of data labeling and model training during technology implementation within a limited budget.
This study first detects the key features in the image by training an instance segmentation model on our Steelscapes dataset which involves two crucial categories: intersection and spacing. In the initial precast U-girder test set, A segmentation AP of 31.006%, an absolute error of 0.166 cm and a relative error of 1.388% are achieved. In addition, two experimental environments were taken as target data pools in accordance with the attributes of manual clustering in order to analyze the test performance of the model in different scenarios. The training results indicate that changing the test situation of the rebar assembly has severe consequences. Therefore, six query strategies for object detection were used to improve the deep learning model performance on the target test set, namely the entropy estimation of three different aggregation methods and the three disagreement measurement methods based on Monte Carlo (MC) dropout. The RoI matching method, which uses average KL divergence to estimate disagreement, has the best efficiency performance in both experiments, reaching the experimental benchmark performance of 81.790% and 92.269% respectively within the fixed labeling budget. Finally, the active learning algorithm combined with manual clustering provides a standard labeling and training process in the continuous training strategy to maximize the utility of the new data.

並列關鍵字

Rebar spacing inspection ； Instance segmentation ； Active learning ； Deep learning ； RGB-D image

參考文獻

Beatty, G., Kochis, E., Bloodgood, M. (2019). The Use of Unlabeled Data Versus Labeled Data for Stopping Active Learning for Text Classification. 2019 IEEE 13th International Conference on Semantic Computing (ICSC), pp. 287-294.

Brust, C. A., Käding, C., Denzler, J. (2018). Active learning for deep object detection. arXiv preprint arXiv:1809.09875.

Canny, J. (1986). A Computational Approach to Edge Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6), pp. 679-698. doi:10.1109/TPAMI.1986.4767851

Gal, Y., Ghahramani, Z. (2016). Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Proceedings of The 33rd International Conference on Machine Learning, 48, pp. 1050-1059. New York, USA.

Girshick, R. (2015). Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1440-1448. doi:10.1109/ICCV.2015.169