改进BiSeNetV1实时模型的岩屑图像识别

在图像分割识别领域,现有的深度学习方法大多使用高精度语义分割方法来实现,存在着网络推理速度慢、计算量大、难以实际应用等问题.借助于表现较好的BiSeNetV1实时网络模型,通过扩展的空间路径卷积结构、空间金字塔注意力机制(SPARM)和简化的注意力特征融合模块(S-iAFF)等改进策略,设计一种用于岩屑图像分割领域的BiSeNet＿SPARM＿S-iAFF实时网络.扩展的空间路径卷积结构可以获取更丰富的岩屑图像空间特征,上下文路径使用优化的空间金字塔注意力机制(SPARM)进一步细化高层语义特征提取,在特征融合阶段使用简化注意力特征融合(S-iAFF)加强低层空间与高层语义特征的融合程度.实验结果表明, BiSeNet＿SPARM＿S-iAFF网络在RockCuttings＿Oil岩屑数据集上的平均交并比(mIoU)为64.91%,相较于BiSeNetV1网络提高了2.68%;另外改进后的网络在精度上接近大部分高精度语义分割方法,同时参数量大幅度减少、推理速度有着明显的提升.

Recognition of Cuttings Images Based on Improved BiSeNetV1 Real-time Model

In the field of image segmentation and identification, the existing deep learning methods mostly perform tasks by high-precision semantic segmentation methods, which lead to a slow network inference speed, large amount of calculation, and difficult actual application. A real-time network model with better performance, namely BiSeNetV1 is used, and the extended spatial path convolution structure, spatial pyramid attention mechanism (SPARM), simplified iterative attention feature fusion (S-iAFF) module, and other optimization strategies are applied. As a result, a real-time BiSeNet_SPARM_S-iAFF network is designed for rock debris image segmentation. The extended spatial path convolution structure can obtain more abundant spatial features of rock debris images. The context path uses the optimized SPARM to further refine high-level semantic feature extraction. Finally, S-iAFF is used to enhance the fusion degree between low-level spatial and high-level semantic features in the feature fusion stage. The experimental results indicate that the mean intersection over union (mIoU) of the BiSeNet_SPARM_S-iAFF network on the RockCuttings_Oil dataset is 64.91%, which is 2.68% higher than that of the BiSeNetV1 network, and the precision of the improved network is close to that of the most high-precision semantic segmentation methods, while the number of parameters is greatly reduced, and the inference speed is significantly improved.