一种超轻量级指静脉纹络实时分割网络

An Ultra-lightweight Real-time Segmentation Network of Finger Vein Textures

现有的指静脉分割网络大多需要消耗极大内存和计算资源,难以直接部署到嵌入式平台上,大部分模型轻量化方法存在参数减小导致分割性能急剧下降、算力受限和实时性等问题。针对上述问题,本文提出了一种超轻量级指静脉纹络实时分割网络—SGUnet。首先,使用沙漏状的深度可分离卷积极大地减少基础模型参数,并采用轻量级高效注意力模块实现无降维的局部跨通道交互,提升网络分割性能。其次,为了解决部分特征图存在冗余的问题,使用Cheap operation来替代部分“懈怠”的卷积核,得到相似的特征图。最后,采用特征信息交互的方法,打开分组卷积的组间通道,解决了分组特征组之间信息不流通的问题。与传统Unet分割网络相比,最终的SGUnet模型参数量约为传统Unet分割网络的1%,Mult-Adds约为0.5%。在两个公开的手指静脉数据集SDU-FV、MMCBNU-6000上验证网络性能,结果表明SGUnet网络在分割性能上不仅优于大型分割网络Unet、DU-Net、R2U-Net,而且超越了经典轻量级改进模型squeeze-Unet、Mobile-Unet、shuffle-Unet、Ghost-Unet。SGUnet网络Accuracy、Dice、AUC分别达到94.11%、0.5384、0.9354,并且在NVIDIA嵌入式平台上指静脉纹络提取的测试速度高达0.27秒/张。

Among biometric recognition technologies,finger vein recognition has attracted the attention of many researchers because of various advantages,such as noncontact collection,living body recognition,forgery difficulty,and low cost.The finger vein extraction is the key step of finger vein recognition technology,which directly affects the accuracy of the finger vein feature extraction,matching,and recognition.Most of the existing finger vein segmentation networks consume considerable memory and computing resources,and deploying them directly to the embedded platform is difficult.The design of lightweight deep neural network architecture is the key to solving this problem.However,most lightweight models have problems,such as sharp decline of segmentation performance,limited computing power,and realtime issue et al.To solve the above problems,this paper proposes an ultralight weight real-time segmentation network of finger vein textures-SGUnet.The SGUnet network realizes real-time finger vein texture extraction on an embedded platform,which is called finger vein segmentation.Moreover,there is a need to comprehensively consider the segmentation performance,network parameter size,and running time.First,the encoding-decoding structure is adopted in the overall network,and the hourglass shaped deep separable volume is used to actively reduce basic model parameters to realize the preliminary lightweight of the model.The lightweight and efficient attention module is used to realize the local crosschannel interaction without dimensionality reduction,improve network segmentation performance,and solve the problem of performance degradation during model compression.The attention module uses a onedimensional convolution neural network to weight the channel in the operation process,while the introduced parameters of the attention module have little effect on the model’s burden.Second,most convolutional neural networks have a feature graph redundancy phenomenon.These redundant feature graphs have great similarities.They can be obtained from similar feature graphs through some simple changes.To solve the problem of partial feature graph redundancy,a swap operation is used to replace some“slack”convolution cores.A similar feature map is obtained through a simple mapping transformation,which ensures the consistency of network output,reduces the part of the convolution kernel,and realizes the second step lightweight of the model.Finally,to further reduce the number of parameters of the channel convolution and the problem that each group of information in group convolution cannot flow,the characteristic information of each group is randomly disrupted and reorganized using the method of characteristic information interaction to realize the information flow between group convolution,further compress the network,and ensure the performance of the model.After the above three steps of lightweight operation,an ultralightweight real-time segmentation network of finger vein textures is finally obtained.To verify the efficiency and real-time performance of this algorithm,two public finger vein databases are used:SDU-FV of Shandong University and MMCBNU-6000 of Quanbei National University of Korea.In the training process,four-fifths of the dataset is randomly selected as the training set and the remaining one-fifth as the test set.In the training and testing,the blocking strategy is adopted for the original image.Each image is divided into multiple patches.When the width and height are five steps,multiple continuous overlapping blocks are extracted from each image.The probability that the pixel is a vein is obtained by averaging the probability of all prediction blocks covering the pixel.To ensure that the memory limit and real-time performance of the hardware platform are not exceeded,selecting the patch with a step of five in terms of index and time is appropriate.After the network outputs the patch results,according to the order of sub patches,the overlapping sliding window strategy is adopted to retain the central region results,discard inaccurate image edges,and resplice them into a complete original image.In the experiment,SGUnet is compared with different segmentation networks,and the comparative experiment is conducted on the embedded platform.Compared with the traditional Unet segmentation network,the parameters of SGUnet model are approximately 1%,and MultAdds are approximately 0.5%of the traditional Unet segmentation network.We verify the network performance on two public finger vein datasets:SDU-FV and MMCBNU-6000.The results show that the segmentation performance of SGUnet network is not only better than that of large segmentation networks Unet,DU-Net,and R2U-net,but also surpasses the classic lightweight models squeeze-Unet,mobile-Unet,shuffle-Unet,and Ghost-Unet,Its performance indexes accuracy,dice and AUC reach 94.11%,0.5384,and 0.9354,respectively.Compared with previous work,the proposed network has made great progress,in which the final parameter is only145K and Flops is only 13M,and it surpasses previous lightweight models.Moreover,SGUnet network meets the low computing power requirements of the embedded platform and can be easily deployed on the whole series of NVIDIA embedded platforms to realize the real-time segmentation of finger vein veins.The test speed of finger vein veins extraction is as high as 0.27 seconds/piece.