Design of multipath lightweight deep network

Abstract

Robot vision is an essential research field that enables robots to perform various tasks by recognizing objects as humans do. Deep convolutional neural networks (DCNNs) have shown many promising results in recent years, there are still critical obstacles that cause degradation in their performance: increase in the number of trainable parameters and computational complexity. Moreover, the classification accuracy of the machine learning algorithms already exceeds that of a well-trained human, and the results are rather saturated. As the applications of deep learning networks become more complex, the size of the model increased rapidly. Nevertheless, deep learning networks are being deployed to lightweight devices such as mobile devices and automobiles. Hence, in recent years, many studies have been conducted in the direction of reducing the weight of the model and applying it to mobile devices. Lightweight model design is designing a model with fewer parameters and computations while maintaining a similar level of performance. If the amount of computations is reduced, it enables deployment of the DCNN on low-power devices and secure real-time performance, and if the number of parameters is reduced, resources required for model storage and transmission are reduced. Therefore, it is very valuable to research lightweight model design. For this purpose, we propose a multipath lightweight deep network using randomly selected dilated convolutions. The model we refer to as the basic structure is DenseNet (including CondenseNet, CondenseNetV2) based because the information of the previous layer is concatenated and the features are reused. The main differences between the proposed network and other network architectures are the existence of multipath networks and the large receptive field effect. The proposed network consists of two sets of multipath networks (minimum 2, maximum 8), where the output feature maps of one path are concatenated with the input feature maps of the other path so that the features are reusable and abundant. We also replace the 3 × 3 standard convolution of each path with a randomly selected dilated convolution, which has the effect of increasing the receptive field. The proposed network lowers the number of floating point operations (FLOPs) and parameters by more than 50% and the classification error by 0.8% as compared to the state-of-the-art. We show that the proposed network is efficient.