Analytic and VLSI Specific Design of Robust CNN Templates

Parameter learning or design is a key issue in cellular neural network (CNN) theory. If the CNN is implemented as an analog VLSI chip, additional constraints are posed due to its restricted accuracy. Only robust parameters will still guarantee the correct network behavior. We present an analytical design approach for the class of bipolar CNNs which yields optimally robust template parameters. We give a rigorous definition of absolute and relative robustness and show that all well-defined CNN tasks are characterized by a finite set of linear and homogeneous inequalities. This system of inequalities can be analytically solved for the most robust template by simple matrix algebra. Focusing on a particular implementation of the CNN universal chip, we demonstrate that the proposed method can cope with the manufacturing inaccuracies.     

New horizon for CNN: with fuzzy paradigms for multimedia

The cellular neural network (CNN) is a powerful technique to mimic the local function of biological neural circuits for real-time image and video processing. Recently, it is widely accepted that using a set of CNNs in parallel can achieve higher-level information processing and reasoning functions either from application or biology points of views. The authors introduce a novel framework for constructing a multiple-CNN integrated neural system called recurrent fuzzy CNN (RFCNN). This system can automatically learn its proper network structure and parameters simultaneously. In the RFCNN, each learned fuzzy rule corresponds to a CNN. Hence, each CNN takes care of a fuzzily separated problem region, and the functions of all CNNs are integrated through the fuzzy inference mechanism. Some online clustering algorithms are introduced for the structure learning, and the ordered-derivative calculus is applied to derive the recurrent learning rules of CNN templates in the parameter-learning phase. RFCNN provides a solution to the current dilemma on the decision of templates and/or fuzzy rules in the existing integrated (fuzzy) CNN systems. The capability of the RFCNN is demonstrated on the real-world vision-based defect inspection and image descreening problems proving that the RFCNN scheme is effective and promising.     

CNN-Based Hybrid-Order Texture Segregation as Early Vision Processing and Its Implementation on CNN-UM

In this paper, a biologically inspired, CNN-based, multi-channel, texture boundary detection technique is presented. The proposed approach is similar to human vision system. The algorithm is simple and straightforward such that it can be implemented on the cellular neural networks (CNNs). CNN contains several important advantages, such as efficient real-time processing capability and feasible very large-scale integration (VLSI) implementation. The proposed algorithm also had been widely tested on synthetic texture images. Those texture images are randomly selected from the Brodatz textures database (1966). According to our simulation results, the boundaries of uniform textures can be detected quite successfully. For the nonuniform or nonregular textures, the results also indicate meaningful properties, and the properties also are consistent to the human visual sensation. The proposed algorithm also has been implemented on the CNN universal machine (CNN-UM), and yields similar results as the simulation on the PC. Based on the efficient performance of CNN-UM, the algorithm becomes very fast.     

A Dedicated Multi-Chip Programmable System for Cellular Neural Networks

Cellular Neural Networks (CNN's) represent a remarkable improvement in the hardware implementation of Artificial Neural Networks (ANN's). In fact, their regular structure and their local connectivity feature contribute to render this class of neural networks especially appealing for VLSI implementations. CNNs are widely applied in several fields, including image processing and pattern recognition. In this research, the authors already presented two fully digitally programmable CNN chips with 3×3 (3×3DPCNN chip) and 6×6 cells (6×6DPCNN chip) respectively. In this paper, a system with twenty of the latter chips will be presented. The main features of this electronic system consist of the full digital programmability of the templates, the digital input/output for logic operations, the analog outputs for dynamic analysis and the implementation of space-variant as well as space-invariant CNNs.     

Aligned and non-aligned double JPEG detection using convolutional neural networks

Due to the wide diffusion of JPEG coding standard, the image forensic community has devoted significant attention to the development of double JPEG (DJPEG) compression detectors through the years. The ability of detecting whether an image has been compressed twice provides paramount information toward image authenticity assessment. Given the trend recently gained by convolutional neural networks (CNN) in many computer vision tasks, in this paper we propose to use CNNs for aligned and non-aligned double JPEG compression detection. In particular, we explore the capability of CNNs to capture DJPEG artifacts directly from images. Results show that the proposed CNN-based detectors achieve good performance even with small size images (i.e., 64 × 64), outperforming state-of-the-art solutions, especially in the non-aligned case. Besides, good results are also achieved in the commonly-recognized challenging case in which the first quality factor is larger than the second one.     

Cellular neural networks: a paradigm for nonlinear spatio-temporalprocessing

The paradigm of Cellular Neural Networks (CNNs)is going to achieve a complete maturity. In fact, from a methodological point of view, important results on their digitally programmable analog dynamics have been gained, completed with thousands of application routines. This has encouraged the spreading of a great number of applications in the most different disciplines. Moreover, their structure, tailor made for VLSI realization, has led to the production of some chip prototypes that, embedded in a computational infrastructure, have produced the first analogic cellular computers. This completes the framework and makes it possible to realize complex spatio-temporal and filtering tasks on a time scale of microseconds. In this paper some sketches on the main aspects of CNNs, from the formal to the hardware prototype point of view, are presented together with some appealing applications to illustrate complex image, visual and spatio-temporal dynamics processing     

A Recurrent Fuzzy Coupled Cellular Neural Network System With Automatic Structure and Template Learning

The cellular neural network (CNN) is a powerful technique to mimic the local function of biological neural circuits, especially the human visual pathway system, for real-time image and video processing. Recently, many studies show that an integrated CNN system can solve more complex high-level intelligent problems. In this brief, we extend our previously proposed multi-CNN integrated system, called recurrent fuzzy CNN (RFCNN) which considers uncoupled CNNs only, to automatically learn the proper network structure and parameters simultaneously of coupled CNNs, which is called recurrent fuzzy coupled CNN (RFCCNN). The proposed RFCCNN provides a solution to the current dilemma on the decision of templates and/or fuzzy rules in the existing integrated (fuzzy) CNN systems. For comparison, the capability of the proposed RFCCNN is demonstrated on the same defect inspection problems. Simulation results show that the proposed RFCCNN outperforms the RFCNN.     

基于细胞神经网络的从阴影恢复形状的新方法

细胞神经网络(CNN)是一种实时处理信号的大规模非线性模拟电路,它的连续时间特点以及局部互连特点使其可以进行并行计算,并且非常适用于超大规模集成电路(VLSI)的实现.本文针对从阴影恢复形状(SFS)问题,提出了一种基于硬件退火CNN的能量函数优化方法,并对该方法进行了详细分析,给出了实例的仿真结果,验证了该方法的有效性.该方法为并行处理算法,具有运算量小、易于大规模VLSI集成实现,且能够克服局部极小等优点,可以使SFS问题得到实时的处理.     

离散时间点格神经网络的开关电流电路实现

本文将原连续时间电压域点格神经网络（CNN）模型[1]映射到离散时间电流域，并采用开关电流（SI）技术来进行离散时间CNN的VLSI设计。给出了基本积木块的实现电路，并进行了连通片检测方面的简单应用研究。计算机仿真与理论结果相吻合，所得电路非常适合于用标准CMOS工艺集成，具有潜在的应用前景。     

Time multiplexed color image processing based on a CNN withcell-state outputs

A practical system approach for time-multiplexing cellular neural network (CNN) implementations suitable for processing large and complex images using small CNN arrays is presented. For real size applications, due to hardware limitations, it is impossible to have a one-on-one mapping between the CNN hardware cells and all the pixels in the image involved. This paper presents a practical solution by processing the input image, block by block, with the number of pixels in a block being the same as the number of CNN cells in the array. Furthermore, unlike other implementations in which the output is observed at the hard-limiting block, the very large scale integrated (VLSI) architecture hereby described monitors the outputs from the state node. While previous implementations are mostly suitable for black and white applications because of the thresholded outputs, our approach is especially suitable for applications in color (gray) image processing due to the analog nature of the state node. Experimental complementary metal-oxide-semiconductor (CMOS) chip results in good agreement with theoretical results are presented     

Analog VLSI Design Constraints of Programmable Cellular Neural Networks

Analog parallel signal processing systems, like cellular neural networks (CNN's), intrinsically have a high potential for perception-like signal processing tasks. The robust design of analog VLSI requires a good understanding of the capabilities as well as the limitations of analog signal processing. Implementation-oriented theoretical methods are described to compute the effect of all types circuit non-idealities with random or systematic causes on the static and dynamical behavior of CNN's and to derive specifications for the cell circuit building blocks. The fundamental impact of transistor mismatch on the trade-off between the speed, accuracy and power performance of CNN chips is demonstrated. A design methodology taking into account the effect of transistor mismatch is proposed and experimental results of a CNN chip implementation designed with this method are discussed.     

A recurrent fuzzy cellular neural network system with automatic structure and template learning

It is widely accepted that using a set of cellular neural networks (CNNs) in parallel can achieve higher level information processing and reasoning functions either from application or biologics points of views. Such an integrated CNN system can solve more complex intelligent problems. In this paper, we propose a novel framework for automatically constructing a multiple-CNN integrated neural system in the form of a recurrent fuzzy neural network. This system, called recurrent fuzzy CNN (RFCNN), can automatically learn its proper network structure and parameters simultaneously. The structure learning includes the fuzzy division of the problem domain and the creation of fuzzy rules and CNNs. The parameter learning includes the tuning of fuzzy membership functions and CNN templates. In the RFCNN, each learned fuzzy rule corresponds to a CNN. Hence, each CNN takes care of a fuzzily separated problem region, and the functions of all CNNs are integrated through the fuzzy inference mechanism. A new online adaptive independent component analysis mixture-model technique is proposed for the structure learning of RFCNN, and the ordered-derivative calculus is applied to derive the recurrent learning rules of CNN templates in the parameter-learning phase. The proposed RFCNN provides a solution to the current dilemma on the decision of templates and/or fuzzy rules in the existing integrated (fuzzy) CNN systems. The capability of the proposed RFCNN is demonstrated on the real-world defect inspection problems. Experimental results show that the proposed scheme is effective and promising.     

Analog image recognition arrays design by using co-fabricated MOSFET and MESFETs on a 0.25?μm SOS process

This paper presents an analog image recognition system with a novel MESFET device fabricated on a fully depleted (FD) CMOS process. An analog image recognition system with a power consumption of 2.4?mW/cell and a settling time of 6.5???s was designed, fabricated and characterized. A CNN is employed to realize a core cell of the proposed image recognition system. While a CNN benefits from its regular structure, it faces challenges due to its power consumption, speed, and size in their CMOS implementations. SOS MESFETs can deal with the challenges associated with CMOS-based CNNs. Advantages of SOS MESFETs associated with nonlinear signal processing include lower power consumption and higher operating speeds compared to similar geometry MOSFETs carrying the same current. SOS MESFET-based analog image recognition systems were fabricated and the transient response is characterized in both simulation using a TOM3 SPICE model extracted from SOS MESFETs and in experiment using image testing lab equipment. Settling times of 3.5 and 6.5???s for one-by-four and one-by-eight arrays, respectively, were achieved with line recognition template. The corresponding power consumption for the two arrays was 9.6 and 19.2?mW, respectively.     

A novel MobileNet with selective depth multiplier to compromise complexity and accuracy

In the last few years, convolutional neural networks (CNNs) have demonstrated good performance while solving various computer vision problems. However, since CNNs exhibit high computational complexity, signal processing is performed on the server side. To reduce the computational complexity of CNNs for edge computing, a lightweight algorithm, such as a MobileNet, is proposed. Although MobileNet is lighter than other CNN models, it commonly achieves lower classification accuracy. Hence, to find a balance between complexity and accuracy, additional hyperparameters for adjusting the size of the model have recently been proposed. However, significantly increasing the number of parameters makes models dense and unsuitable for devices with limited computational resources. In this study, we propose a novel MobileNet architecture, in which the number of parameters is adaptively increased according to the importance of feature maps. We show that our proposed network achieves better classification accuracy with fewer parameters than the conventional MobileNet.     

Tunneling-Based Cellular Nonlinear Network Architectures for Image Processing

The resonant tunneling diode (RTD) has found numerous applications in high-speed digital and analog circuits due to the key advantages associated with its folded back negative differential resistance (NDR) current-voltage (I-V) characteristics as well as its extremely small switching capacitance. Recently, the RTD has also been employed to implement high-speed and compact cellular neural/nonlinear networks (CNNs) by exploiting its quantum tunneling induced nonlinearity and symmetrical I-V characteristics for both positive and negative voltages applied across the anode and cathode terminals of the RTD. This paper proposes an RTD-based CNN architecture and investigates its operation through driving-point-plot analysis, stability and settling time study, and circuit simulation. Full-array simulation of a 128 $,times,$128 RTD-based CNN for several image processing functions is performed using the Quantum Spice simulator designed at the University of Michigan, where the RTD is represented in SPICE simulator by a physics based model derived by solving SchrÖdinger's and Poisson's equations self-consistently. A comparative study between different CNN implementations reveals that the RTD-based CNN can be designed superior to conventional CMOS technologies in terms of integration density, operating speed, and functionality.      

A 6 × 6 Cells Interconnection-Oriented Programmable Chip for CNN

The implementation of a versatile VLSI chip certainly represents an important step to improve the research on Cellular Neural Networks. In this paper a VLSI realization of the multi-chip oriented, the 6 × 6 Digitally Programmable Cellular Neural Network (6 × 6 DPCNN) chip, will be presented. This chip covers most of the available one-neighbourhood templates for image processing applications. Moreover, it can be easily interconnected to others to carry out very large CNN arrays. The designs and some measured results of a single chip and a multi-chip board (the 720 DPCNN System) will be shown.     

Signal processing using cellular neural networks

The cellular neural network (CNN) architecture combines the best features from traditional fully-connected analog neural networks and digital cellular automata. The network can rapidly process continuous-valued (gray-scale) input signals (such as images) and perform many computation functions which traditionally were implemented in digital form. Here, we briefly introduce the the theory of CNN circuits, provide some examples of CNN applications to image processing, and discuss work toward a CNN implementation in custom CMOS VLSI. The role of analog computer-aided design (CAD) will be briefly presented as it relates to analog neural network implementation.This work is supported in part by the Office of Naval Research under Contract N00014-89-J1402, and the National Science Foundation under grant MIP-8912639.     

Optimization and regularization of complex task decomposition for blind removal of multi-factor degradation

Most existing image restoration methods based on deep neural networks are developed for images which only degraded by a single degradation mode and imaging under an ideal condition. They cannot be directly used to restore the images degraded by multi-factor coupling. A complex task decomposition regularization optimization strategy (TDROS) is proposed to solve the problem. The restoration of images degraded by multi-factor coupling is a complex task that can be solved by separating these multiple factors, that is, breaking the complex task into numbers of simpler tasks to make the entire complex problem be overcome more easily. Motivated by this idea, the TDROS decomposes the complex task of image restoration into two sub-task: the potential task constrained by regularization and the main task for reconstructing high-definition images. In TDROS, the front of the neural network is focused on the restoration of images degraded by additive noise, while the other part of the network is focused mainly on the restoration of images degraded by blur. We applied the TDROS to an 11-layer convolutional neural network (CNN) and compared it with initial CNNs from the aspects of restoration accuracy and generalization ability. Based on these results, we used TDROS to design a novel network model for the restoration of atmospheric turbulence-degraded images. The experimental results demonstrate that the proposed TDROS can improve the generalization ability of the existing network more effectively than current popular methods, offering a better solution for the problem of severely degraded image restoration. Moreover, the TDROS concept provides a flexible framework for low-level visual complex tasks and can be easily incorporated into existing CNNs.     

PAC-Bayesian framework based drop-path method for 2D discriminative convolutional network pruning

Deep convolutional neural networks (CNNs) have demonstrated its extraordinary power on various visual tasks like object detection and classification. However, it is still challenging to deploy state-of-the-art models into real-world applications, such as autonomous vehicles, due to their expensive computation costs. In this paper, to accelerate the network inference, we introduce a novel pruning method named Drop-path to reduce model parameters of 2D deep CNNs. Given a trained deep CNN, pruning paths with different lengths is achieved by ordering the influence of neurons in each layer on the probably approximately correct (PAC) Bayesian boundary of the model. We believe that the invariance of PAC-Bayesian boundary is an important factor to guarantee the generalization ability of deep CNN under the condition of optimizing as much as possible. To the best of our knowledge, this is the first time to reduce model size based on the generalization error boundary. After pruning, we observe that the convolutional kernels themselves become sparse, rather than some being removed directly. In fact, Drop-path is generic and can be well generalized on multi-layer and multi-branch models, since parameter ranking criterion can be applied to any kind of layer and the importance scores can still be propagated. Finally, Drop-path is evaluated on two image classification benchmark datasets (ImageNet and CIFAR-10) with multiple deep CNN models, including AlexNet, VGG-16, GoogLeNet, and ResNet-34/50/56/110. Experimental results demonstrate that Drop-path achieves significant model compression and acceleration with negligible accuracy loss.

     

Modular Cellular Neural Network Structure for Wave‐Computing‐Based Image Processing

This paper introduces the modular cellular neural network (CNN), which is a new CNN structure constructed from nine one‐layer modules with intercellular interactions between different modules. The new network is suitable for implementing many image processing operations. Inputting an image into the modules results in nine outputs. The topographic characteristic of the cell interactions allows the outputs to introduce new properties for image processing tasks. The stability of the system is proven and the performance is evaluated in several image processing applications. Experiment results on texture segmentation show the power of the proposed structure. The performance of the structure in a real edge detection application using the Berkeley dataset BSDS300 is also evaluated.