An algorithm was proposed on the base of on an improved SparseInst network to address the low accuracy and poor real-time performance of road crack detection algorithms. The algorithm improved fine crack segmentation by introducing SPM stripe pooling mod-ules and MPM mixed pooling modules, and enhanced crack recognition completeness and accuracy by incorporating CBAM modules and DCNv2 to capture road crack characteristics. Meanwhile, Gaussian blur and OTSU thresholding segmentation methods were used to re-duce image noise to increase the difference between background and cracks, and morphological methods were used to skeletonize the crack image width to a unit pixel value. Experimental results show that, compared to the original SparseInst network, this algorithm im-proves AP, AP50 , AP75 by 3. 5％, 0. 8％, and 2. 1％, respectively, with measurement errors of 2. 3％ ～13. 3％, demonstrating its high effectiveness and practicality.
维普
万方数据