| Summary: | 碩士 === 國立中央大學 === 資訊工程學系 === 107 === In recent years, face verification has been widely used to secure various transactions on the internet. The current state-of-the-art in face verification is convolutional neural network (CNN). Despite the performance of CNN, deploying CNN in mobile and embedded devices is still challenging because the available computational resource on these devices is constrained. In this paper, we propose a lightweight CNN for face verification using several methods. First, a modified version of ShuffleNet V2 called ShuffleHalf is used as the backbone network for the FaceNet algorithm. Second, the feature maps in the model are reused using two proposed methods called Reuse Later and Reuse ShuffleBlock. Reuse Later works by reusing the potentially unused features by connecting the features directly to the fully connected layer. Meanwhile, Reuse ShuffleBlock works by reusing the feature maps output of the first 1x1 convolution in the basic building block of ShuffleNet V2 (ShuffleBlock). This method is used to reduce the percentage of 1x1 convolution in the model because 1x1 convolution operation is computationally expensive. Third, kernel size is increased as the number of channels increases to obtain the same receptive field size with less computational complexity. Fourth, the depthwise convolution operations are used to replace some ShuffleBlocks. Fifth, other existing previous state-of-the-art algorithms are combined with the proposed method to see if they can increase the performance-efficiency tradeoff of the proposed method.
Experimental results on five testing datasets show that ShuffleHalf achieves better accuracy than all other baselines with only 48% FLOPs of the previous state-of-the-art algorithm, MobileFaceNet. The accuracy of ShuffleHalf is further improved by reusing the feature. This method can also reduce the computational complexity to only 42% FLOPs of MobileFaceNet. Meanwhile, both changing kernel size and using depthwise repetition can further decrease computational complexity to only 38% FLOPs of MobileFaceNet with better performance than MobileFaceNet. Combination with some existing methods does not increase the accuracy nor performance-efficiency tradeoff of the model. However, adding shortcut connections and using Swish activation function can improve the accuracy of the model without any noticeable increase in the computational complexity.
|
|---|
