基于改进U-Net网络的混凝土表面裂缝分割

如何快速、高效、准确地像素级分割混凝土表面裂缝是当前研究的热点问题之一.在混凝土表面裂缝图像中裂缝面积远远小于正常路面面积,造成现有方法在这种正负样本分布不均问题中无法有效学习裂缝特征,最终分割效果较差.提出了一种将Focal损失与活动轮廓模型相结合的新损失函数,针对裂缝面积较小且连续分布的特点,通过Focal损失加强...

Crack segmentation of concrete surface based on improved U-Net

How to detect and segment concrete surface cracks quickly, efficiently and accurately at pixel level has been one of the focused areas in current research. In the image of cracks on the concrete surface, the crack area is much smaller than the normal road area, which makes the existing method unable to effectively learn from the crack feature in the imbalanced distribution of positive and negative samples. This finally leads to the poor performance of existing models. In this paper, a novel loss function is proposed, which combines the focal loss and the active contour model. For the small and continuous characteristics of the crack area, Focal loss can enhance the sensitivity of the model to the crack. Active contour model can guarantee that the segmentation result is consistent with the real result. Batch normalization layer is added to the model''s convolution block to enhance the activation effect and suppress the generation of gradient oscillation during the model training. To deploy the model in an embedded environment such as a vehicle detector, we add pruning and quantization while ensuring the segmentation accuracy to achieve the compression model size. The experimental results show that the proposed method can effectively study crack characteristics and accurately and efficiently identify cracks.