Bridge Damage Detection Based on Improved SSD Algorithm

For the task of bridge damage detection, the bridge damage object detection dataset is proposed based on the dacl10k bridge damage semantic segmentation dataset. To further enhance the real-time performance of detection, the extra feature layers in the backbone of the SSD network is substituted with a depthwise separable convolution from the general convolution, and the original backbone of the SSD network is replaced with a lightweight MobileNetV2 network to boost the inference speed while ensuring accuracy. Since bridge damage data tends to be lengthy, and the prior bounding box aspect ratio in the original SSD network is configured based on the general dataset, the K-means algorithm is employed to cluster the prior bounding box aspect ratio that aligns with the characteristics of the dataset, aiming to discover the optimal initial value for the model. This approach expedites the convergence speed and augments the network accuracy. The improved algorithm reduces the number of parameters to 25% of the original one, while the precision is still similar to the previous one.