Computational Neural Network Research Articles

Booth Encoding-Based Energy Efficient Multipliers for Deep Learning Systems

Artificial intelligence on edge is a growing research field. In this paper, we propose a novel re-encoding scheme for reducing the size of the weights of deep neural networks (DNNs). The proposed re-encoding scheme exploits the Booth encoding scheme and the power-of-two (PO2) quantization to allow for very low energy computations during the inference of the neural networks with minimal loss in classification accuracy. We demonstrate the advantages of the proposed re-encoding scheme by computing a convolutional neural network (CNN) and a linear neural network on the proposed Extended Exact Multiplier and the proposed PO2 Multiplier. Our proposed PO2 quantization and re-encoding method reduce the model size for the CNN by 30.77% and the model size of the linear neural network by 49.86%. Furthermore, our multipliers reduce the inference energy for CNN by 50.6% and for the linear neural network by 90.1%. The PO2 Multiplier is proposed for the sensor-end computation of the linear neural network with a 77.32% reduction in the area relative to an exact Booth multiplier and it reduces the inference energy consumption of the linear neural network by 93.2% when compared to the unmodified exact multiplier. Our proposed scheme can be used to improve the energy consumption during inference for most Booth multipliers with minor modifications to the re-encoding signal arrangements. We also demonstrate that the proposed re-encoding scheme paired with the proposed multipliers outperforms all the existing designs in terms of resource utilization with a minimal impact on the inference accuracy of the neural networks.

Comparing representations and computations in single neurons versus neural networks

Single-neuron-level explanations have been the gold standard in neuroscience for decades. Recently, however, neural-network-level explanations have become increasingly popular. This increase in popularity is driven by the fact that the analysis of neural networks can solve problems that cannot be addressed by analyzing neurons independently. In this opinion article, I argue that while both frameworks employ the same general logic to link physical and mental phenomena, in many cases the neural network framework provides better explanatory objects to understand representations and computations related to mental phenomena. I discuss what constitutes a mechanistic explanation in neural systems, provide examples, and conclude by highlighting a number of the challenges and considerations associated with the use of analyses of neural networks to study brain function.

Progress on Memristor-Based Analog Logic Operation

There is always a need for low-power, area-efficient VLSI (Very Large-Scale Integration) design and this need is increasing day by day. However, conventional design methods based on Metal-Oxide-Semiconductor Field Effect Transistor (MOSFET) devices and Complementary Metal-Oxide-Semiconductor Transistor (CMOS) technology cannot meet the performance requirements. The memristor, as a promising computing and memory integration device, offers a new research idea for conventional logic circuit structure and architecture innovation, given its non-volatility, scalability, low power consumption, fast switching speed, etc. This paper proposes a brief overview of the characteristics and current status of memristor-based logic circuits and analyzes their applications in numerical expression and memory. The benefits and drawbacks of various analog logic circuit structures are summarized and compared. In addition, some solution strategies for these issues are presented. Finally, this paper offers prospects for the applications of memristors in the logic implementation of large-scale memristor arrays, the novel structure of in-memory computing, and neural network computing.

Three levels of information processing in the brain

Information, the measure of order in a complex system, is the opposite of entropy, the measure of chaos and disorder. We can distinguish several levels at which information is processed in the brain. The first one is the level of serial molecular genetic processes, similar in some aspects to digital computations (DC). At the same time, higher cognitive activity is probably based on parallel neural network computations (NNC). The advantage of neural networks is their intrinsic ability to learn, adapting their parameters to specific tasks and to external data. However, there seems to be a third level of information processing as well, which involves subjective consciousness and its units, so called qualia. They are difficult to study experimentally, and the very fact of their existence is hard to explain within the framework of modern physical theory. Here I propose a way to consider consciousness as the extension of basic physical laws – namely, total entropy dissipation leading to a system simplification. At the level of subjective consciousness, the brain seems to convert information embodied by neural activity to a more simple and compact form, internally observed as qualia. Whereas physical implementations of both DC and NNC are essentially approximate and probabilistic, qualia-associated computations (QAC) make the brain capable of recognizing general laws and relationships. While elaborating a behavioral program, the conscious brain does not act blindly or gropingly but according to the very meaning of such general laws, which gives it an advantage compared to any artificial intelligence system.

Compact lensless convolution processor for an optoelectronic convolutional neural network

To our knowledge, optical 4f systems have been widely used as a convolutional layer to perform convolutional computation in free-space optical neural networks (ONNs), which makes ONNs too bulky to be easily applied to miniaturized smart systems. Hence, we propose a compact lensless optoelectronic convolutional neural network (LOE-CNN) architecture in which a single optimized diffractive phase mask acts as an analog convolution processor to perform convolutional operation without a Fourier lens or lenslet array. We demonstrate that this LOE-CNN can be functionally comparable to existing electronic counterparts in classification performance, achieving a classification accuracy of 98.07% and 95% over the Modified National Institute of Standards and Technology dataset in simulation and experiment, respectively, which not only opens new application prospects for free-space ONNs based on a compact single-chip convolution processor, but also facilitates the development of ONN-based smart devices.

Detection of Bad Stapled Nails in Wooden Packages

Wooden nail-stitched crates are widely used for fruit transportation. Bad stapled nails are transformed into severe product damage that creates stains on the crate due to its juice. In consequence, the final customer depreciates the product because the quality product is in doubt. Human visual inspection of badly stapled nails is a non-effective solution since constant criteria are difficult to reach for all of crate production. A system for the in-line inspection based on a conveyor belt of badly stapled nails in stitched crates is presented. The developed inspection system is discussed with the definition of the computer vision system used to identify fails and the description of image processing algorithms. The experiments are focused on a comparative analysis of the performance of five state-of-the-art classification algorithms based on a deep neural network and traditional computer vision algorithms, highlighting the trade-off between speed and precision in the detection. An accuracy of over 95% is achieved if the user defines the nail location in the image. The presented work constitutes a benchmark to guide deep-learning computer vision algorithms in realistic applications.

Neuronal motifs reveal backbone structure and influential neurons of neural network in C. elegans

Abstract Neural network elements such as motif, backbone and influential nodes play important roles in neural network computation. Increasing researches have been applying complex network methods in order to identify different essential structures within complex neural networks. However, the distinct properties of synapses that build the neural network are often neglected, such as the difference between chemical synapses and electrical synapses. By separating these distinct synapses, we can identify a novel repertoire of neural motifs and greatly expand neural motif families in neural systems. Based on the expanded motif families, we further propose a novel neural-motif-based algorithm to extract the backbone in the neural network. The backbone circuit we extracted from Caenorhabditis elegans connectome controls an essential motor behaviour in C. elegans. Furthermore, we develop a novel neural-motif-based algorithm to identify influential neurons. Compared with the influential neurons identified using existing methods, the neurons identified in this work provide more information in related to their functions. These methods have been successfully applied to identify a series of network features in C. elegans, providing a biologically interpretable way of exploring the structure of neural network.

Thermal instability analysis in magneto-hybrid nanofluid layer between rough surfaces with variable gravity and space-dependent heat source

The onset of thermal instability with hybrid nanoparticles Al2O3–Cu in nanofluid is investigated with the combined effects of variable gravity, variable heat source and Lorentz force in a porous medium. Its applications are in many areas like chemical engineering, geophysics, astrophysics, etc. Based on literature, many gravity variations are assumed in the present analysis, along with a space-dependent heat source/sink parameter that varies along the width of the channel. The finite difference-based Lobatto IIIa method has been applied to solve the system under no nanoparticle flux and rough boundary conditions, and approximate analytical results are generated for some special cases. The roughness parameters ([Formula: see text]), Darcy number (Da), gravity variation function ([Formula: see text]), and Chandrasekhar number (Q) delay the onset of convection and thus stabilize the system. It is also observed that an increase in the Lewis number Le, power index in variable heat source, and the nanoparticle Rayleigh concentration number Rn decreases the critical Rayleigh number [Formula: see text] which destabilizes the system, and increases the critical wave number, which enlarges the convection cell size. In the case of exponential variation ([Formula: see text]) in the gravity variation parameter, the system becomes stabilized due to a delay in the onset of convection. In addition, we have considered multilayer perceptron-artificial neural network (MLP-ANN) computation to predict the critical Rayleigh number as per function of important controlling parameters. The data set of 625 observations is chosen keeping 70% for testing, 15% for training and 15% for validation using efficient Levenberg−Marquardt back propagation algorithm with optimal accuracy measures i.e., root mean square deviation (RMSE), root mean relative error (RMRE) and [Formula: see text] (coefficient of determination). Finally, the regression plots are drawn that correlate, target and output data.

Gradient free stochastic training of ANNs, with local approximation in partitions

We present a numerical scheme for computation of Artificial Neural Networks (ANN) weights, which stems from the Universal Approximation Theorem, avoiding costly iterations. The proposed algorithm adheres to the underlying theory, is highly fast, and results in remarkably low errors when applied to regression and classification problems of complex data sets with {textbf{x}} in {mathbb {R}}^{n} (e.g. Griewank, Gomez-Levy, Shekel, and Polynomial functions) with random noise addition (i.e. Uniform, Normal, Generalized Pareto, Log-Normal, and a mixture of Log-Normal, Exponential, and Frechet), as well as the database for handwritten digits recognition MNIST (Modified National Institute of Standards and Technology) with 7times 10^4 images. The same mathematical formulation was found capable of approximating highly nonlinear functions in multiple dimensions, with low errors (e.g. 10^{-10}) for the test set of the unknown functions, their higher-order partial derivatives, as well as numerically solving Partial Differential Equations, such as those appearing in Physics, Engineering, Environmental Sciences, etc. The method is based on the calculation of the weights of each neuron in small neighbourhoods of the data. Accordingly, optimization of hyperparameters is not necessary, as the number of neurons stems directly from the dimensionality of the data, further improving the algorithmic speed. Under this setting, overfitting is inherently avoided, and the results are interpretable and reproducible. The complexity of the proposed algorithm is of class P with {mathcal {O}}(mNni_{cl} + Nmn^2+Nn^3 + mN^2+N^3) computing time, with respect to the observations m, features n, and Neurons N, contrary to the NP-Complete class of standard algorithms for ANN training. The performance of the method is high, irrespective of the size of the data set, and the test set errors are similar or smaller than the training errors, indicating the generalization efficiency of the algorithm. A supplementary computer code in Julia and Python Languages is provided, which can be used to reproduce the validation examples, and/or apply the algorithm to other data sets.

Abstract P5-02-33: Deep learning-based assessment of HER2-low expression on breast cancer H&E digital whole slide images

Abstract Background: Antibody drug conjugates (ADCs) against HER2 have shown meaningful clinical activity in HER2 low breast cancers, defined as 1+ or 2+ staining on immunohistochemistry (IHC) without gene amplification by in situ hybridization (ISH) techniques. Given that these methods were originally developed for an accurate detection of HER2 3+, their sensitivity and robustness for the detection of low and ultra-low levels of HER2 are questionable. We have recently described a deep learning algorithm that can detect signatures of HER2 expression based on training utilizing scanned H&E whole slide images (WSI) of breast cancers for which IHC and mRNA expression levels of HER2 were available. Here, we report the application of our algorithm to two independent breast cancer cohorts. Methods: A model was developed based on recognition of invasive breast cancer in whole slide images of H&E staining, and then trained via computational neural network with multiple instance learning for binary classification of cases as HER2 “negative” and HER2 “expressed” (low). For training, true negatives were defined as having HER2 IHC-0 and mRNA level < 7.6. HER2-low cases were defined as IHC-1+/2+ and mRNA >9. IHC-0 cases with mRNA >7.6 were excluded from the training cohorts. The resulting model (HER2Complete) was able to distinguish HER2-negatives from HER2-low cases with an AUC of 0.91 (+/- 0.08). Here we use Her2Complete to assess HER2 in two additional cohorts that include 901 ER+/HER2 IHC-0 and 52 HER2 IHC 0+ breast cancers from MSK and TCGA cohorts, respectively. For the TCGA cohort, concomitant transcriptomics data (RNASeq) as a reference for HER2 mRNA expression were retrieved and “HER2 expressed” defined as RNASeq expression of HER2 greater than the 90th percentile of the geometric mean of expression of three reference genes not expressed in breast tissues (TTN, MUC13, OR10A6). Values less than this reference cut-off in the TCGA cohort were considered “HER2 not expressed.” Results: Among the 901 IHC-0 test cases from the MSK cohort, the model identified 82 as ‘negative’, whereas 819 were found to have features of HER2 expression (HER2-Low). Of the 82 negative cases in the MSK cohort, all except 13 cases expressed mRNA levels < 9, and 786/819 of the HER2-low cases expressed mRNA levels >8. Of the 52 IHC 0+ cases in the TCGA cohort, 33 also had “HER2 not expressed” by our reference based RNASeq expression cut-off. Our model identified 15 of these 33 as ‘negative’, while 15 of the 19 TCGA cases with IHC 0+ and HER2 ‘expressed’ by our cut-off were identified as ‘HER2-Low’ by our model. Conclusions: AI tools based on the analysis of WSIs of routinely prepared H&E sections may predict HER2 status in breast cancer. This work requires further investigation using treatment response data to demonstrate that cases with morphologic features of low level HER2 expression will respond to ADCs. Citation Format: Gerard Oakley III, Jorge Reis-Filho, David Klimstra, Marc Goldfinger, Yikan Wang, Antonio Marra. Deep learning-based assessment of HER2-low expression on breast cancer H&E digital whole slide images [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P5-02-33.

DyNNamic: Dynamically Reshaping, High Data-Reuse Accelerator for Compact DNNs

Convolutional layers dominate the computation and energy costs of Deep Neural Network (DNN) inference. Recent algorithmic works attempt to reduce these bottlenecks via compact DNN structures and model compression. Likewise, state-of-the-art accelerator designs leverage spatiotemporal characteristics of convolutional layers to reduce data movement overhead and improve throughput. Although both are independently effective at reducing latency and energy costs, combining these approaches does not guarantee cumulative improvements due to inefficient mapping. This inefficiency can be attributed to (1) inflexibility of underlying hardware and (2) inherent reduction of data-reuse opportunities of compact DNN structures. To address these issues, we propose a dynamically reshaping, high data-reuse PE array accelerator, namely <i>DyNNamic</i> . DyNNamic leverages kernel-wise filter decomposition to partition the convolution operation into two compact stages: Shared Kernels Convolution (SKC) and Weighted Accumulation (WA). Because both stages have vastly different dimensions, DyNNamic reshapes its PE array to effectively map the algorithm to the architecture. The architecture then exploits data-reuse opportunities created by the SKC stage, further reducing data movement with negligible overhead. We evaluate our approach on various representative networks and compare against state-of-the-art accelerators. On average, DyNNamic outperforms DianNao by <inline-formula><tex-math notation="LaTeX">$8.4\times$</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">$12.3\times$</tex-math></inline-formula> in terms of inference energy and latency, respectively.

An Efficient Deep-Learning-Based Super-Resolution Accelerating SoC With Heterogeneous Accelerating and Hierarchical Cache

This article presents an energy-efficient accelerating system-on-chip (SoC) for super-resolution (SR) image reconstruction on a mobile platform. With the rise of contactless communication and streaming services, the need for SR is growing. As one of the most basic low-level image processing algorithms, SR can reconstruct high-quality images from low-quality images which are noisy, compressed, or with damaged pixels. However, a massive amount of computation and considerable precision of pixel data pose challenges for acceleration in a resource and bandwidth constrained platform. SR has high energy consumption and long latency. While previous neural processing units (NPUs) reduced the precision to increase the efficiency and accelerate convolutional neural network (CNN) computation, few of them concentrated on both the output image quality and the performance of the entire system. The proposed SR SoC restores the high-quality image using a precision-optimized SR algorithm on an energy-efficient accelerating architecture and cache subsystem. It contributes three algorithm-hardware co-optimized features: 1) heterogeneous accelerating architecture (HAA) with only 8-bit floating-point (FP)-and-fixed-point (FXP) hybrid-precision for SR task; 2) tile-based hierarchical cache (THC) subsystem for the low energy and small footprint cost layer fusion; and 3) heterogeneous L1 data lifetime-aware optimized cache (DLOC) for the energy-efficient on-chip memory access. The prototype of SR SoC is fabricated in 65-nm technology and occupies a 10.0-mm2 die area. The proposed SR SoC can maintain the high reconstruction quality while consuming only 19% of the energy of an FXP16 system with homogeneous NPU. As a result, the SR SoC presents <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2.6\times $ </tex-math></inline-formula> higher energy efficiency than the previous SR targeting NPU and achieves 107-frame-per-second (fps) framerates running <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4\times $ </tex-math></inline-formula> SR image generation to full high definition (FHD) scale at only 0.92-mJ/frame energy consumption.

Fast Computational Approach to the Levenberg-Marquardt Algorithm for Training Feedforward Neural Networks

Abstract This paper presents a parallel approach to the Levenberg-Marquardt algorithm (LM). The use of the Levenberg-Marquardt algorithm to train neural networks is associated with significant computational complexity, and thus computation time. As a result, when the neural network has a big number of weights, the algorithm becomes practically ineffective. This article presents a new parallel approach to the computations in Levenberg-Marquardt neural network learning algorithm. The proposed solution is based on vector instructions to effectively reduce the high computational time of this algorithm. The new approach was tested on several examples involving the problems of classification and function approximation, and next it was compared with a classical computational method. The article presents in detail the idea of parallel neural network computations and shows the obtained acceleration for different problems.

Hyperspectral Image Classification with Deep CNN Using an Enhanced Elephant Herding Optimization for Updating Hyper-Parameters

Deep learning approaches based on convolutional neural networks (CNNs) have recently achieved success in computer vision, demonstrating significant superiority in the domain of image processing. For hyperspectral image (HSI) classification, convolutional neural networks are an efficient option. Hyperspectral image classification approaches are often based on spectral information. Convolutional neural networks are used for image classification in order to achieve greater performance. The complex computation in convolutional neural networks requires hyper-parameters that attain high accuracy outputs, and this process needs more computational time and effort. Following up on the proposed technique, a bio-inspired metaheuristic strategy based on an enhanced form of elephant herding optimization is proposed in this research paper. It allows one to automatically search for and target the suitable values of convolutional neural network hyper-parameters. To design an automatic system for hyperspectral image classification, the enhanced elephant herding optimization (EEHO) with the AdaBound optimizer is implemented for the tuning and updating of the hyper-parameters of convolutional neural networks (CNN–EEHO–AdaBound). The validation of the convolutional network hyper-parameters should produce a highly accurate response of high-accuracy outputs in order to achieve high-level accuracy in HSI classification, and this process takes a significant amount of processing time. The experiments are carried out on benchmark datasets (Indian Pines and Salinas) for evaluation. The proposed methodology outperforms state-of-the-art methods in a performance comparative analysis, with the findings proving its effectiveness. The results show the improved accuracy of HSI classification by optimising and tuning the hyper-parameters.

Neural network estimation of kinetic parameters in distributed activation energy model (DAEM) without a priori assumptions for parallel reaction system

In this study, a new estimation method for the kinetic parameters in a distributed activation energy model (DAEM) was designed and developed. In the proposed method, the conversion estimation by the DAEM is regarded as a feedforward computation of a three-layer neural network, and the kinetic parameters of the DAEM are estimated by optimization of the neural network. The proposed method does not require an a priori assumption of the kinetic parameters or mechanism of a parallel reaction system. First, we created reaction data using numerical simulations, and a kinetic analysis using the neural network was performed. The neural network predicted conversion X very accurately; however, reactions with low contributions to the parallel reaction system also appeared. Next, we carried out a kinetic analysis using the neural network with the lower limit on the contribution of the ith reaction Vi*/V*. The lower limit on Vi*/V* did not influence the prediction accuracy of X and had a significant effect on reducing the reactions with low Vi*/V* and the estimation accuracies of the kinetic parameters in the DAEM. The optimal value of the lower limit on Vi*/V* was determined to be 1.0×10−5–1.0×10−4 when using the neural network with 64 hidden layer nodes. Moreover, the prediction accuracy of the proposed method was compared with those of conventional methods — the single-Gaussian method (1-DAEM), double-Gaussian method (2-DAEM), and Miura and Maki method. The kinetic parameters estimated using the proposed method were closer to the true values than those obtained using conventional methods. Moreover, the X value predicted by the proposed method was more accurate than that predicted by conventional methods. The influence of approximation on training data creation was also examined. The estimation accuracy of the neural network was still high but slightly deteriorated when the neural network was optimized using the reaction data created without the approximation.

Toward a cerebello-thalamo-cortical computational model of spinocerebellar ataxia

Computational neural network modelling is an emerging approach for optimization of drug treatment of neurological disorders and fine-tuning of rehabilitation strategies. In the current study, we constructed a cerebello-thalamo-cortical computational neural network model to simulate a mouse model of cerebellar ataxia (pcd5J mice) by manipulating cerebellar bursts through reduction of GABAergic inhibitory input. Cerebellar output neurons were projected to the thalamus and bidirectionally connected with the cortical network. Our results showed that reduction of inhibitory input in the cerebellum orchestrated the cortical local field potential (LFP) dynamics to generate specific motor outputs of oscillations of the theta, alpha, and beta bands in the computational model as well as in mouse motor cortical neurons. The therapeutic potential of deep brain stimulation (DBS) was tested in the computational model by increasing the sensory input to restore cortical output. Ataxia mice showed normalization of the motor cortex LFP after cerebellum DBS. We provide a novel approach to computational modelling to investigate the effect of DBS by mimicking cerebellar ataxia involving degeneration of Purkinje cells. Simulated neural activity coincides with findings from neural recordings of ataxia mice. Our computational model could thus represent cerebellar pathologies and provide insight into how to improve disease symptoms by restoring neuronal electrophysiological properties using DBS.

Emergent computations in trained artificial neural networks and real brains

Synaptic plasticity allows cortical circuits to learn new tasks and to adapt to changing environments. How do cortical circuits use plasticity to acquire functions such as decision-making or working memory? Neurons are connected in complex ways, forming recurrent neural networks, and learning modifies the strength of their connections. Moreover, neurons communicate emitting brief discrete electric signals. Here we describe how to train recurrent neural networks in tasks like those used to train animals in neuroscience laboratories and how computations emerge in the trained networks. Surprisingly, artificial networks and real brains can use similar computational strategies.

AINNS: All-Inclusive Neural Network Scheduling Via Accelerator Formalization

Driven by the rapid development of accelerators and diverse efficiency requirements of the naturally heterogeneous neural network computation, recent years have seen increased heterogeneity in neural network accelerator systems in terms of network structures, accelerator dataflows and implementations. However, existing research fails to schedule and map the heterogeneous neural networks on heterogeneous accelerators efficiently. They rely on clumpy exhaustive search or complicated ad hoc mapping approaches due to the semantic gap between the networks and accelerators. This paper proposes a systematic method to transform various accelerators into standard parameterized containers of the neural network loops, which builds a direct connection between the computation and the underlying hardware resources. This enables us to match the neural networks with accelerators based on their essential characteristics (e.g., reuse opportunities and bandwidth requirements) without diving into the detailed architectures. To this end, we propose AINNS, an all-inclusive neural network scheduler, that automatically schedules and maps the NN computation on heterogeneous accelerators with just one universal algorithm. Our experimental results show the proposed AINNS not only performs well in the traditional neural network acceleration but also improves the system throughput and energy efficiency by 1.8x and 1.7x respectively in the most challenging heterogeneous acceleration system.

Hand Gesture Recognition

Abstract: This work presents a computer-vision-based application for recognizing hand gestures. A live video feed is captured by a camera, and a still image is extracted from that feed with the aid of an interface. At least once per count hand gesture (one, two, three, four, and five), the system is trained. After that, the system is given a test gesture to see if it can identify it. Several algorithms that are capable of distinguishing a hand gesture were studied. It was determined that the highest rate of accuracy was achieved by using the computational neural network known as the Alexnet algorithm. Traditionally, systems have used data gloves or markers as a means of input. We are free to use the system however we like. In this way, the user can make natural hand gestures in front of the camera. The system implemented serves as an extendable basis for future work toward a fully robust hand gesture recognition system, which is still the subject of intensive research and development.

VFR: The Underwater Acoustic Target Recognition Using Cross-Domain Pre-Training with FBank Fusion Features

Underwater acoustic target recognition is a hot research area in acoustic signal processing. With the development of deep learning, feature extraction and neural network computation have become two major steps of recognition. Due to the complexity of the marine environment, traditional feature extraction cannot express the characteristics of the targets well. In this paper, we propose an underwater acoustic target recognition approach named VFR. VFR adopts a novel feature extraction method by fusing three-dimensional FBank features, and inputs the extracted features into a residual network, instead of the classical CNN network, plus cross-domain pre-training to perform target recognition. The experimental results show that VFR achieves 98.5% recognition accuracy on the randomly divided ShipsEar dataset and 93.8% on the time-divided dataset, respectively, which are better than state-of-the-art results.