應用研究 2024年 第53卷 第21期
DOI: 10. 7672 / sgjs2024210006
基于無人機與深度學習的少樣本混凝土表面裂縫檢測方法
張慧樂1,楊發1,吳丹1,張淳杰2
作者簡介：
張慧樂，高級工程師，E?mail:zhanghuile@ cribc. com作者單位：
1.中國京冶工程技術有限公司，北京 100088; 2.北京交通大學，北京 100089基金項目：
?國家自然科學基金（62072026）摘要：
混凝土表面裂縫檢測是混凝土建筑安全評估和風險預警的重要手段。傳統人工檢測方法工作量大，且需考慮復雜環境下影響人身安全等因素。基于無人機的裂縫檢測方法得到了應用，但當受不確定因素影響時，無人機無法采集足夠的訓練樣本，限制了其檢測性能。為此，基于無人機與深度學習，提出少樣本條件下混凝土表面裂縫檢測方法，采用主流深度學習網絡 Faster?RCNN 和 YOLOX,利用 WBF 算法將檢測結果進行融合，有效彌補了像素級標簽信息較少導致的檢測性能下降。在少樣本裂縫圖像庫及戶外場地進行了試驗測試，試驗結果表明，在少樣本條件下基于無人機與深度學習的裂縫檢測方法性能得到有效提升，對裂縫檢測的準確率達到 58. 67%.
English：
Concrete surface crack detection is an important means of concrete building safety assessmentand risk earlywarning. The traditional manual detection method has a large workload, and it is necessaryto consider factors such as personal safety in complex environments. The crack detection method based onUAV has been applied, but when affected by uncertain factors, UAV cannot collect enough trainingsamples,which limits their detection performance. Therefore, based on UAV and deep learning, a crackdetection method for concrete surface under the condition of few samples is proposed. The mainstreamdeep networks Faster?RCNN and YOLOX are used, and the detection results are fused by WBFalgorithm,which effectively alleviates the detection performance degradation caused by less pixel?levellabel information. Experimental tests were carried out in a few sample crack image library and outdoorsites. The test results show that the performance of the crack detection method based on UAV and deeplearning is effectively improved under the condition of few samples, and the accuracy of crack detection is58. 67%.