Over the past decade, deep-learning-based representations have demonstrated remarkable performance in academia and industry. The learning capability of convolutional neural networks (CNNs) originates from a combination of various feature extraction layers that fully utilize a large amount of data. However, they often require substantial computation and memory resources while replacing traditional hand-engineered features in existing systems. In this review, to improve the efficiency of deep learning research, we focus on three aspects: quantized/binarized models, optimized architectures, and resource-constrained systems. Recent advances in light-weight deep learning models and network architecture search (NAS) algorithms are reviewed, starting with simplified layers and efficient convolution and including new architectural design and optimization. In addition, several practical applications of efficient CNNs have been investigated using various types of hardware architectures and platforms.
31 October 2022 in Journal of Institute of Control, Robotics and Systems
In this paper, we present a lightweight deep learning model for room layout estimation. We build on the work in HorizonNet, in which a 3D room is restored from a single panoramic picture using three steps (pre-processing, feature extraction, and post-processing) by studying the efficient computation of the feature extraction part. In contrast to the baseline method that uses a combination of a typical residual network (ResNet) and long short-term memory (LSTM) as its principal architecture, we focus on the use of a platform-aware neural architecture search for mobile applications (MnasNet) and a gated recurrent unit (GRU) instead of conventional LSTM. Subsequently, hyperparameters of the suggested architectures are selected using sampling-based optimization. In our qualitative and quantitative experiments, the lightweight model using the combination of MnasNet and GRU required approximately half as many parameters compared to the original method for competitive performance of room layout estimation. Based on the Manhattan world assumption, the proposed architecture was validated using Stanford2D3D, PanoContext, and our real-world panorama dataset collected by off-the-shelf software in RICOH THETA Z1.