A Study of Image Recognition for Standard Convolution and Depthwise Separable Convolution

Abstract

Artificial intelligence and deep learning techniques are all around our life. Image recognition and natural language processing are the two major topics. Through using TensorFlow-GPU as backend in convolutional neural network (CNN) and deep learning network, image recognition has been an extreme breakthrough in recent years. However, more and more model parameters result in overfitting problem and computation overhead. In the paper, the performance of image recognition between standard CNN and depthwise separable CNN is experimented and investigated. In addition, data augmentation technique is applied to both standard and depthwise separable CNNs to improve the image recognition accuracy. The experiments are implemented by an open source API called Keras with using CIFAR-10 dataset. Experimental results showed that the depthwise separable CNN improves validation accuracy compared with the standard CNN. Moreover, schemes with data augmentation achieve higher validation accuracy but training accuracy.