“…Mohammad et al [12] implemented a demodulation method based on a deep convolutional neural network (DCNN) and compared its performance to those of other machine learning and non-learning methods for demodulation of a Rayleigh-faded wireless data signal with several settings of signal-to-noise ratio (SNR). They showed that DCNN was able to achieve a lower bit error rate than other methods in all experimental scenarios.…”
Section: Brief Reviewmentioning
confidence: 99%
“…Furthermore, for MiniDenseNet, we maintain the architecture of DenseNet but apply a smaller building block block_config = (2,2,4,4) and growth rate k = 16 for it. This is the smallest model with the lowest complexity (see Table 2) among the others, especially when compared to the original DenseNet121 architecture of which building block and growth rate are block_config = (6,12,24,16) and k = 64, respectively. In the case of unfixed LED positions, an STN layer is required for the automatic transformation of input images.…”
We have proposed a square wave quadrature amplitude modulation (SW-QAM) scheme for visible light communication (VLC) using an image sensor in our previous work. Here, we propose a robust and unified system by using a neural decoding method. This method offers essential SW-QAM decoding capabilities, such as LED localization, light interference elimination, and unknown parameter estimation, bundled into a single neural network model. This work makes use of a convolutional neural network (CNN) that has a capability in automatic learning of unknown parameters, especially when it deals with images as an input. The neural decoding method can provide good solutions for two difficult conditions that are not covered by our previous SW-QAM scheme: unfixed LED positions and multiple point spread functions (PSFs) of multiple LEDs. Responding to the above solutions, three recent CNN architectures-VGG, ResNet, and DenseNet-are modified to suit our scheme and other two small CNN architectures-VGG-like and MiniDenseNet-are proposed for low computing devices. Our experimental results show that the proposed neural decoding method performs better in terms of error rate than the theoretical decoding, an SW-QAM decoder with a Wiener filter, in different scenarios. Furthermore, we experiment on the problem of moving camera, i.e., the unfixed position of LED points. For this case, a spatial transformer network (STN) layer is added to the neural decoding method for solving the moving camera problem, and the method with the new layer achieves a remarkable result. INDEX TERMS Visible light communication, image sensor communication (ISC), SW-QAM, optical camera communication (OCC), neural decoding, convolutional neural network (CNN), deep learning.
“…Mohammad et al [12] implemented a demodulation method based on a deep convolutional neural network (DCNN) and compared its performance to those of other machine learning and non-learning methods for demodulation of a Rayleigh-faded wireless data signal with several settings of signal-to-noise ratio (SNR). They showed that DCNN was able to achieve a lower bit error rate than other methods in all experimental scenarios.…”
Section: Brief Reviewmentioning
confidence: 99%
“…Furthermore, for MiniDenseNet, we maintain the architecture of DenseNet but apply a smaller building block block_config = (2,2,4,4) and growth rate k = 16 for it. This is the smallest model with the lowest complexity (see Table 2) among the others, especially when compared to the original DenseNet121 architecture of which building block and growth rate are block_config = (6,12,24,16) and k = 64, respectively. In the case of unfixed LED positions, an STN layer is required for the automatic transformation of input images.…”
We have proposed a square wave quadrature amplitude modulation (SW-QAM) scheme for visible light communication (VLC) using an image sensor in our previous work. Here, we propose a robust and unified system by using a neural decoding method. This method offers essential SW-QAM decoding capabilities, such as LED localization, light interference elimination, and unknown parameter estimation, bundled into a single neural network model. This work makes use of a convolutional neural network (CNN) that has a capability in automatic learning of unknown parameters, especially when it deals with images as an input. The neural decoding method can provide good solutions for two difficult conditions that are not covered by our previous SW-QAM scheme: unfixed LED positions and multiple point spread functions (PSFs) of multiple LEDs. Responding to the above solutions, three recent CNN architectures-VGG, ResNet, and DenseNet-are modified to suit our scheme and other two small CNN architectures-VGG-like and MiniDenseNet-are proposed for low computing devices. Our experimental results show that the proposed neural decoding method performs better in terms of error rate than the theoretical decoding, an SW-QAM decoder with a Wiener filter, in different scenarios. Furthermore, we experiment on the problem of moving camera, i.e., the unfixed position of LED points. For this case, a spatial transformer network (STN) layer is added to the neural decoding method for solving the moving camera problem, and the method with the new layer achieves a remarkable result. INDEX TERMS Visible light communication, image sensor communication (ISC), SW-QAM, optical camera communication (OCC), neural decoding, convolutional neural network (CNN), deep learning.
“…In demodulation, a generalized bit level demodulation scheme for M-ary QAM systems is proposed in [10], which significantly reduces the complexity and has almost the same bit error rate (BER) performance as the ML algorithm. The authors in [11] propose the deep convolutional neural network to demodulate the Rayleigh-faded signal and the results show the deep convolutional neural network has a lower bit error probability compared to other demodulators such as the support vector machine. In [12], the authors propose the deeplearning-based demodulator in short-range multipath channels, where the deep belief network and the stacked autoencoder are applied to their demodulation system.…”
Over the past few decades, the information theory community has worked to develop modulation and encoding that achieve the Shannon capacity in the constraint of the low implementation complexity. In this paper, we focus on the demodulation/decoding of the complex modulations/codes that approach the Shannon capacity. Theoretically, the maximum likelihood (ML) algorithm can achieve the optimal error performance whereas it has O(2 k ) demodulation/decoding complexity with k denoting the number of information bits. Recent progress in deep learning provides a new direction to tackle the demodulation and the decoding. The purpose of this paper is to analyze the feasibility of the neural network to demodulate/decode the complex modulations/codes close to the Shannon capacity and characterize the error performance and the complexity of the neural network. Regarding the neural network demodulator, we use the golden angle modulation (GAM), a promising modulation format that can offer the Shannon capacity approaching performance, to evaluate the demodulator. It is observed that the neural network demodulator can get a close performance to the ML-based method while it suffers from the lower complexity order in the low-order GAM. Regarding the neural network decoder, we use the Gaussian codebook, achieving the Shannon capacity, to evaluate the decoder. We also observe that the neural network decoder achieves the error performance close to the ML decoder with a much lower complexity order in the small Gaussian codebook. Limited by the current training resources, we cannot evaluate the performance of the high-order modulation and the long codeword. But, based on the results of the low-order GAM and the small Gaussian codebook, we boldly give our conjecture: the neural network demodulator/decoder is a strong candidate approach for demodulating/decoding the complex modulations/codes close to the Shannon capacity owing to the error performance of the near-ML algorithm and the lower complexity.
“…In [5] a convolutional neural network (CNN) was used to demodulate the bipolar extended binary phase shift keying signal to solve the problem of serious inter-symbol interference. In [6] CNN was used to realize FSK demodulation under Rayleigh fading channels. All of these works considered the problem of hard demodulation.…”
Soft demodulation is a basic module of traditional communication receivers. It converts received symbols into soft bits, that is, log likelihood ratios (LLRs). However, in the nonideal additive white Gaussian noise (AWGN) channel, it is difficult to accurately calculate the LLR. In this letter, we propose a demodulator, DemodNet, based on a fully convolutional neural network with variable input and output length. We use hard bit information to train the DemodNet, and we propose log probability ratio (LPR) based on the output layer of the trained DemodNet to realize soft demodulation. The simulation results show that under the AWGN channel, the performance of both hard demodulation and soft demodulation of DemodNet is very close to the traditional methods. In three non-ideal channel scenarios, i.e., the presence of frequency deviation, additive generalized Gaussian noise (AGGN) channel, and Rayleigh fading channel, the performance of channel decoding using the soft information LPR obtained by DemodNet is better than the performance of decoding using the exact LLR calculated under the ideal AWGN assumption.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.