AM-ResNet: Low-energy-consumption addition-multiplication hybrid ResNet for pest recognition

Abstract

Pest recognition technology has rapidly progressed in a short period with the development of deep convolutional neural networks. However, the vast calculation burden of these networks requires massive energy, especially for specific applications such as all-day-working pest monitoring systems, which process images from dozens of devices uninterruptedly. This paper proposes a low-energy-consumption hybrid ResNet structure – AM-ResNet, consisting of addition-based and multiplication-based convolutional layers to address this problem. This paper presents an optimal AM-ResNet design method through a detailed experimental analysis of the performance differences between building blocks in two typical ResNet variants, ResNet20 and ResNet32. Then, this method is applied to construct a deep AM-ResNet for pest recognition, which significantly reduces the energy consumption at the cost of an acceptable accuracy loss. Experiments show that the proposed network performs well in our proposed pest dataset PEST20. Extensive experiments demonstrate that the network can save more than 40% energy by losing less than 2% accuracy in IP102 and VOC2007. In addition, by analyzing the visualization results, this paper summarizes the advantages of the two convolutions. It presents the application direction for the hybrid networks.