基于双线性注意力金字塔网络的压疮等级识别

Pressure ulcer grade recognition based on bilinear attention pyramid network

在实际临床应用场景下,压疮伤口等级分类较多且图像间差距小。针对使用图像识别技术对压疮图像分类难度大的问题,提出了基于双线性注意力金字塔的压疮等级识别网络BAP-CNN。该网络以细粒度分类网络APCNN为基础(细粒度指类间差距小的情况符合压疮等级分类需求),设计了瓶颈注意力模块,并且采用双线性注意力池化的方法,从而提升模型的整体性能和识别准确率。实验结果表明,在细粒度视觉分类数据集和自建压疮伤口图像SCU-PU数据集上,改进后的网络BAP-CNN与基础网络APCNN以及经典细粒度网络NTS、WSDAN相比,模型的识别率均有所提升,证明了改进方法的有效性,以及在不同数据集下良好的泛化能力。

In actual clinical application scenarios,there are many grades of pressure ulcer wounds and the gap between images is small.Aiming at the difficulty of using image recognition technology to classify pressure ulcer images,a pressure ulcer grade recognition network BAP-CNN based on bilinear attention pyramid is proposed.The network is based on the fine-grained classification network APCNN.The fine-grained means that the small gap between classes meets the requirements of pressure ulcer grade classification.The bottleneck spatial attention module is introduced,and the bilinear attention pooling method is used to improve the model.Therefore,overall performance and recognition accuracy of the model is improved.The experimental results show that on the fine-grained visual classification dataset and the self-built pressure ulcer wound image SCU-PU dataset,compared with the basic network APCNN and the classic fine-grained networks NTS and WSDAN,the recognition of the improved network BAP-CNN is better.The rate of improvement has been improved,which proves the effectiveness of the improved method and the good generalization ability under different datasets.