Small-Sample Object Detection of Surface Cracks in Concrete Structures of High-Rise Buildings via Multi-Level Transfer Learning

To ensure realistic crack effects on the complex surfaces of high-rise concrete structures and meet the demands of small-sample target detection, a small-sample target detection method for surface cracks in high-rise concrete structures is proposed under multi-level transfer learning. A two-dimensional maximum entropy threshold segmentation method is employed to segment images of high-rise concrete structures. After obtaining the target image, crack connectivity area filtering and crack linearity and rectangularity filtering are applied to remove isolated noise points. A multi-level transfer learning architecture is constructed by integrating multi-scale hybrid temporal convolutional networks, long short-term memory neural networks, and Attention mechanisms to generate distinct transfer learning hidden layers. Processed images are input as source domain data into this architecture, enabling knowledge transfer through the generated multi-level hidden layers. After small-sample hierarchical training, shared features between source and target domains are extracted. A cosine classifier outputs the crack category detection results for high-rise concrete structures. Test results demonstrate that this method accurately captures the irregular contours of mesh cracks and effectively distinguishes crack regions from backgrounds. It efficiently removes isolated point noise in images, maintaining smoothness metrics consistently between 0.008 and 0.015. The approach adapts to detecting cracks of diverse morphologies and categories.