水体微纤维图像识别的改进MobileNetV2算法

An Improved MobileNetV2 Algorithm for Image Recognition of Microfibers in Water

针对人工识别水体微纤维耗时耗力,以及传统图像处理算法识别水体微纤维图像鲁棒性弱等问题,构建了一种改进的MobileNetV2网络识别微纤维算法。在特征提取部分采用特征重构策略,先压缩深度卷积特征,获取全局感受野;再利用多层全连接为每个通道生成权重,建立通道之间的相互依赖关系;最后逐通道加权到原特征上,完成对原始特征的重构。此外,采用不同大小的下采样器捕获不同尺度的特征信息并融合,增强微纤维的细节特征信息,提升模型对微纤维的学习能力与识别效果。改进MobileNetV2网络的微纤维识别准确率达到97.96%,与原始MobileNetV2网络相比高2.54%,同时,误识率和漏识率也有显著的降低。相较于ResNet、DenseNet、VGG16和NasNet网络,模型大小压缩了若干倍,微纤维识别准确率有所提升,误识率与漏识率大大降低。实验表明:该网络模型能够提取更加完整的微纤维特征信息,加强微纤维特征判别指向性的同时减小了模型尺寸,降低了在移动设备中部署的难度,并且使识别微纤维具有更高的准确率和更好的稳定性。

Aiming at the problems of time-consuming and labor-consuming manual identification of water microfibers,and the weak robustness of traditional image processing algorithms for identifying water microfiber images,an improved MobileNetV2 network identification method for microfibers is constructed.In the feature extraction part,the feature reconstruction strategy is adopted.Firstly,the deep convolution features are compressed to obtain the global receptive field.Then,the fully connected layers are used to generate weights for each channel to establish the interdependence between the channels.Finally,the channel is weighted to the original in terms of features to complete the reconstruction of the original features.In addition,different sizes of downsamplers are used to capture and fuse feature information of different scales to enhance the detailed feature information of microfibers,and to improve the model′s learning ability and recognition effect of microfibers.The improved MobileNetV2 network′s microfiber recognition accuracy rate reaches 97.96%.Compared with the original MobileNetV2 network,the recognition accuracy rate is increased by 2.54%.At the same time,the false recognition rate and the missed recognition rate are also significantly reduced.In comparison to ResNet,DenseNet,VGG16 and NasNet networks,the model size is compressed several times,the accuracy of microfiber recognition is improved,and the false recognition rate and missed recognition rate are greatly reduced.Experimental results show that the network model can extract more complete feature information for microfiber.While strengthening the microfiber feature to identify the directivity,the model is reduced,and the difficulty of deployment in mobile devices is reduced as well.The improved model recognizes microfibers with higher accuracy and better stability.