Although complicated and deep neural network models can achieve high accuracy of image recognition, they require huge amount of computations and parameters and are not suitable for mobile and embedded devices. As a result, MobileNet was proposed, which can reduce the amount of parameters and computational cost dramatically. In this paper, we propose two different methods to improve MobileNet, which are based on adjusting two hyper-parameters width multiplier and depth multiplier, combing max pooling or Fractional Max Pooling with MobileNet. We tested the improved models on images classification database CIFAR-10 and CIFAR-100 with good results .
View on arXiv