Research on automatic identification method of building construction defects based on UAV image and convolutional neural network

Aiming at the problems of high cost, low efficiency and strong subjectivity of traditional manual inspection, and the shortcomings of existing defect identification methods in complex scene adaptability, fine-grained defect detection and real-time, this study innovatively introduces multi-scale feature fusion, attention mechanism and lightweight model design. In the stage of UAV image acquisition and preprocessing, the three-dimensional modeling of the construction area is realized based on RTK positioning and SLAM technology, and the dark channel prior defogging algorithm and CLAHE algorithm are used for image preprocessing. The multi-scale feature fusion network architecture is based on the improved EfficientNet-B4, and the dynamic hole convolution module (D-DConv) and the improved Feature Pyramid Network (FPN) are introduced, and the dual-path attention module (DPA) is embedded in the FPN to improve the robustness of the model to complex background and the ability to identify small defects. In order to realize the efficient deployment of the model in UAV, a lightweight reasoning optimization strategy is proposed, including teacher-student knowledge distillation framework and dynamic quantification strategy. In addition, the multi-task joint optimization objective function is designed, and the three sub-tasks of classification, location and pixel-level segmentation are considered comprehensively to improve the detection accuracy and robustness. The experimental results show that this method is superior to many benchmark models in mAP(92.4%) and small defect recall rate (89.2%), with the lowest number of parameters (only 4.1M), and achieves better detection performance while maintaining a high reasoning speed (45.0 FPS). The lightweight strategy significantly reduces the model volume and improves the reasoning speed while maintaining high accuracy. The model volume is reduced from 4.1M to 0.71M, and the mAP is only reduced by 1.2%. The real-time reasoning of 45FPS is realized on Jetson Nano platform, which is 3.1 times faster than that of FP32 mode. In the continuous monitoring of a super-high-rise project for 30 days, this method greatly reduces the time spent on a single detection from 6 hours to 25 minutes, achieving a speed increase of 14.4 times, The recognition rate of hidden dangers increased from 68% to 93.5%, an increase of 25.5%, The average false detection rate has also dropped from 23.1 times per thousand images to 4.8 times, a decrease of 79.2%, which is significantly better than the traditional manual inspection method.