基于并行注意力UNet的裂缝检测方法

裂缝对公共设施而言存在着安全隐患，因此裂缝检测是公共设施进行维护的重要手段.由于裂缝图像中存在噪声、光线、阴影等因素干扰，神经网络在训练时极易被影响，导致预测结果出现偏差，降低预测效果.为减少这些干扰，设计了一个并行注意力机制，并将其嵌入到UNet网络的解码部分，进而提出了并行注意力UNet(parallel attention based UNet, PA-UNet).该方法分别从通道和空间2个维度加大裂缝特征权重以抑制干扰，然后对这2个维度生成的特征进行融合，以获得更具互补性的裂缝特征.为了验证该方法的有效性，选取了4个数据集进行实验，结果表明该方法较现有的主流方法，裂缝检测效果更加优异.同时，为了验证并行注意力机制的有效性，选取了4种注意力机制与其进行对比实验，结果表明并行注意力机制效果优于其他注意力机制.

Parallel Attention Based UNet for Crack Detection

Cracks have hidden safety hazards to public facilities, so crack detection is essential for the maintenance of public facilities. Due to the interference of noise, light, shadow, and other factors in the crack images, the neural network is easily affected during the training process, which causes deviations in the prediction results and reduces the prediction effect. To suppress these disturbances, a parallel attention mechanism is designed and then the parallel attention based UNet(PA-UNet) is proposed by embedding this attention mechanism into UNet. The parallel attention mechanism increases the weights of crack features from the two dimensions of channel and space to suppress interference, then fuses the features generated by these two dimensions to obtain more complementary crack features. To verify the effectiveness of the proposed method, we have conducted experiments on four data sets. Experimental results show that our method outperforms the existing popular methods. Meanwhile, to demonstrate the effectiveness of the parallel attention mechanism, we conduct a comparative experiment with other four attention mechanisms. The results show that the parallel attention mechanism performs better than others.