【摘要】 针对现有裂缝检测算法提取的局部梯度信息不足而导致裂缝识别精度低的问题，提出了一种基于像素差分的编码器-解码器U型结构的裂缝检测方法。使用由像素差分卷积块和残差块组成的PDCBlock作为网络的编码器，将两种不同方向上的像素邻域的差分值计算融合到卷积网络运算中，以充分获取裂缝语义信息的同时更好地捕获局部梯度信息，使得在背景复杂的情景下裂缝边缘识别更准确。在混凝土类岩石裂缝数据集CRACK500上对比了同类方法，实验结果表明：该方法在上述数据集上的平均交并比和D_sc分别为0.436 3、0.580 4，裂缝分割精度、相似性上均优于对比方法。更多还原

【Abstract】 Existing crack detection algorithms suffer from low crack recognition accuracy due to insufficien local gradient information.To address the problem,this paper proposes a crack detection method (pixe different U-Net,PDU-Net)based on pixel difference encoder-decoder U-shaped structure.Specifically,we adopted PDCBlock as the encoder of proposed method,which is composed of residual blocks and pixe difference convolution blocks.Then the difference values of pixel neighborhoods in two different directions can be computationally fused into the convolutional network operations.With the help of PDCBlock,our proposed network enables the semantic information of crack to be fully accessible and also better captures the local gradient information,which makes the crack edge recognition more accurate under complex background Similar methods were compared on the concrete-similar rock crack dataset CRACK500.Extensive experiments demonstrate that the proposed method obtains a mean intersection over union of 0.436 3 and an average Dice similarity coefficient of 0.580 4,and outperforms existing state-of-the-art methods in terms of crack segmentation accuracy and similarity on the above data sets.更多还原