Abstract: Loosening and corrosion of communication tower bolts are the key hidden dangers that threaten the safety of infrastructure. However, there is a challenge of extreme class imbalance in the actual detection data set. The normal state is dominant, and the samples of defects such as loosening and corrosion are scarce, which leads to the poor performance of the traditional deep learning model in the safety-sensitive minority class recall rate ( no false negative rate ). In order to solve this problem, this paper proposes a heterogeneous dual-stream convolutional network ( Dual-Stream CNN ) based on depth estimation guidance and its robust loss optimization strategy. Firstly, the model adopts the RGB-Raw Depth dual-stream architecture ( based on ResNet-18 ), which effectively integrates the texture information ( RGB ) and geometric structure features ( Depth ) of the image to enhance the model "s ability to discriminate looseness ( geometric deformation ) and corrosion ( surface texture ). Secondly, for extreme class imbalance, we innovatively combine Focal Loss ( γ = 1.5 ) with Loose class radical over-frequency weight ( × 2.8 ), and achieve a breakthrough improvement in its recall rate by increasing the attention to Loose, the most sparse and critical defect. Finally, in order to make all defect categories meet the industrial safety standards, we introduce a Corroded category decision threshold tuning strategy ( the threshold is reduced to 0.10 ). The dataset in this study consists of 3,886 images, with 3,109 for training and 777 for validation.Experiments were conducted using the PyTorch framework on an NVIDIA RTX 3090 GPU. The experimental results show that after the above systematic optimization, the method achieves an overall accuracy of 96.78 % on the test set, and the recall rate of key defects is fully up to standard : Normal recall rate reaches 97.64 %, Loose recall rate reaches 100.00 %, and Corroded recall rate reaches 90.35 %. This study successfully solves the problem of defect underreporting under extreme imbalance, and provides a robust and efficient technical solution for intelligent maintenance of key infrastructure such as communication towers.