Usando convolução separável em profundidade na otimização da arquitetura SqueezeNet

Great advances have been made in models based on Convolutional Neural Networks. Considering image processing problems, these models already outperform humans in some cases. Despite the excellent results in this area, models with extremely high accuracy rates are, in general, very large, reaching hundreds of millions of parameters, which makes them unfeasible for several real-world applications, where the computational power available is usually quite limited. In this context, we investigated the possible reduction in the number of parameters of the SqueezeNet network, which aims to be an architecture with reduced size and good accuracy, caused by the replacement of its traditional convolutions by Depthwise Separable Convolutions (DSC), as well as the impact of this replacement on other model analysis metrics. The metrics analyzed are the accuracy, the number of parameters, the storage size and the inference time of a single test example. The resulting network, called SqueezeNet-DSC, is then applied to the image classification problem, and its performance is compared to other networks that are a reference in the area, such as MobileNet, AlexNet and VGG19. SqueezeNet-DSC showed a considerable reduction in storage space, reaching 37% of the storage space of the original SqueezeNet, with a loss of accuracy of 1.07% on the CIFAR-10 base and 3.06% on CIFAR-100 database.