基于自注意力语义分割的航空发动机孔探图像检测

Borehole image detection of aero-engine based on self-attention semantic segmentation model

针对传统语义分割模型对于航空发动机孔探图像内损伤的检测存在小尺度或高相似度损伤易被漏检误判的问题,提出了一种基于自注意力语义分割(SA-SS)模型的航空发动机孔探图像检测方法。基于语义分割模型DeepLabv3+的总体架构,采用轻量级MobileNetV2替代原始的Xception作为主干特征提取网络,利用扩张—提取—压缩的结构进行特征提取,以减少模型计算量。基于多层级联结构,改进原始DeepLabv3+的空洞空间金字塔池化结构,使特征图保有更丰富的特征信息。在模型内融合一种自注意力机制,建立全局像素的内部相关性,加强对细节特征的注意力。改进原始DeepLabv3+的解码层,将多尺度空间融合方法引入低层特征提取,融合多个跃层特征。实验结果表明:与传统DeepLabv3+、SegNet-ResNet等方法相比,SA-SS模型的平均交并比和平均像素精确度最大分别提升了4.10%和3.92%,训练时间和平均检测速度最大分别改善了24.43%和5.11帧/s。

Aiming at the problems that small-scale faults tend to be missed and misjudged when detecting borehole images of aero-engines by using traditional methods,a new method based on self-attention semantic segmentation(SA-SS)model is proposed.Based on the overall architecture of classical semantic segmentation model DeepLabv3+,a lightweight MobileNetV2 is adopted as the backbone feature extraction network instead of Xception to reduce calculation by utilizing expansion-extraction-compression strategy;based on the idea of multi-layer cascade,original atrous spatial pyramid pooling structure of DeepLabv3+is improved to keep more feature information in feature map;a self-attention mechanism is fused to establish the internal correlation of global pixels and strengthen the attention to details.The decoding layer of original DeepLabv3+is improved;multi-scale spatial fusion method is introduced into low-level feature extraction to fuse multiple layers of features for classification.Experimental results show that compared with original DeepLabv3+,SegNet-ResNet and other methods,mean intersection over union and pixel accuracy and PA of SA-SS are increased by 4.10%and 3.92% respectively.Also,training cost and detection speed are improved by 24.43% and 5.11frame/s respectively.