Showing papers on "Convolutional neural network published in 2004"

Low resolution OCR for camera acquired documents

Abstract: A global optimization framework for optical character recognition (OCR) of low-resolution photographed documents that combines a binarization-type process, segmentation, and recognition into a single process. The framework includes a machine learning approach trained on a large amount of data. A convolutional neural network can be employed to compute a classification function at multiple positions and take grey-level input which eliminates binarization. The framework utilizes preprocessing, layout analysis, character recognition, and word recognition to output high recognition rates. The framework also employs dynamic programming and language models to arrive at the desired output.

Genetic algorithm optimization of a convolutional neural network for autonomous crack detection

Abstract: Detecting cracks is an important function in building, tunnel and bridge structural analysis. Successful automation of crack detection can provide a uniform and timely means for preventing further damage to structures. This laboratory has successfully applied convolutional neural networks (CNNs) to online crack detection. CNNs represent an interesting method for adaptive image processing and form a link between artificial neural networks, and finite impulse response filters. As with most artificial neural networks, the CNN is susceptible to multiple local minima, thus complexity and time must be applied in order to avoid becoming trapped within the local minima. This paper employs a standard genetic algorithm (GA) to train the weights of a 4-5x5 filter CNN in order to pass through the local minima. This technique resulted in a 92.3/spl plusmn/1.4% average success rate using 25 GA-trained CNNs presented with 100 crack (320x240 pixel) images.

A VLSI convolutional neural network for image recognition using merged/mixed analog-digital architecture

Abstract: Hierarchical convolutional neural networks represent a well-known robust image-recognition model. In order to apply this model to robot vision or various intelligent vision systems, its VLSI implementation with high performance and low power consumption is required. This paper proposes a VLSI convolutional network architecture using a hybrid approach composed of pulse-width modulation (PWM) and digital circuits. We call this approach merged/mixed analog-digital architecture. The VLSI chip includes PWM neuron circuits, PWM/digital converters, digital adder-subtracters, and digital memory. We have designed and fabricated a VLSI chip by using a 0.35 μm CMOS process. The VLSI chip can perform 6-bit precision convolution calculations for an image of 100 × 100 pixels with a receptive field area of up to 20 × 20 pixels within 5 ms, which means a performance of 2 GOPS. Power consumption of PWM neuron circuits was measured to be 20 mW. We have verified successful operations using a fabricated VLSI chip.

A face detection system using shunting inhibitory convolutional neural networks

Abstract: We present a face detection system based on a class of convolutional neural networks, namely shunting inhibitory convolutional neural networks (SICoNNets). The topology of these networks is a flexible feedforward architecture with three different connections schemes: fully-connected, toeplitz-connected and binary-connected. SICoNNets were trained, using a hybrid method based on Rprop, Quickprop and least squares, to discriminate between face and non-face patterns. All three connection schemes achieve 99% detection accuracy at 5% false alarm rate, based on a test set of 7000 face and non-face patterns. Furthermore, toeplitz-connected network was trained on a larger training set and has achieved a 99% correct classification rate with only 1% false alarm rate based on the same test set. A face detection system is built based on the trained convolutional neural networks. The system accepts an input image of arbitrary size and localizes the face patterns in the image. To localize faces of different sizes, the convolutional neural network is applied as a face detection filter at different scales. The detection scores from different scales are aggregated together to form the final decision.

Convolutional Spiking Neural Network for Robust Object Detection with Population Code Using Structured Pulse Packets

Abstract: We propose a convolutional spiking neural network (CSNN) model with population coding for robust object (e.g., face) detection. Basic structure of the network involves hierarchically alternating layers for feature detection and feature pooling. The proposed model implements hierarchical template matching by temporal integration of structured pulse packet. The packet signal represents some intermediate or complex visual feature (e.g., a pair of line segments, corners, eye, nose, etc.) that constitutes a face model. The output pulse of a feature pooling neuron represents some local feature (e.g., end-stop, blob, eye, etc.). Introducing a population coding scheme in the CSNN architecture, we show how the biologically inspired model attains invariance to changes in size and position of face and ensures the efficiency of face detection.

Unsupervised Feature Selection for Multi-class Object Detection Using Convolutional Neural Networks

Abstract: Convolutional Neural Networks (CNN) have proven to be useful tools for object detection and object recognition. They act like feature extractor and classifier at the same time. In this study we present an unsupervised feature selection procedure for constructing a training set for the CNN and analyze in detail the learnt receptive fields. We then introduce, for the first time, a figural alphabet to be used for low-level feature detection with CNN. This alphabet turned out to be useful in detecting a vocabulary set of intermediate level features and considerably reduces the complexity of the CNN. Moreover we propose an optimal high-level feature selection procedure and apply this to the challenging problem of car detection. We demonstrate promising results for multi-class object detection using obtained figural alphabet to detect considerably different categories of objects (e.g., faces and cars).

Face recognition using SVM combined with CNN for face detection

Abstract: We propose a model for face recognition using a support vector machine being fed with a feature vector generated from outputs in several modules in bottom as well as intermediate layers of convolutional neural network (CNN) trained for face detection. The feature vector is composed of a set of local output distributions from feature detecting modules in the face detecting CNN. The set of local areas are automatically selected around facial components (e.g., eyes, moth, nose, etc.) detected by the CNN. Local areas for intermediate level features are defined so that information on spatial arrangement of facial components is implicitly included as output distribution from facial component detecting modules. Results demonstrate highly efficient and robust performance both in face recognition and in detection as well.

Entropy-based generation of supervised neural networks for classification of structured patterns

Abstract: Sperduti and Starita proposed a new type of neural network which consists of generalized recursive neurons for classification of structures. In this paper, we propose an entropy-based approach for constructing such neural networks for classification of acyclic structured patterns. Given a classification problem, the architecture, i.e., the number of hidden layers and the number of neurons in each hidden layer, and all the values of the link weights associated with the corresponding neural network are automatically determined. Experimental results have shown that the networks constructed by our method can have a better performance, with respect to network size, learning speed, or recognition accuracy, than the networks obtained by other methods.

A Convolutional Neural Network VLSI Architecture Using Thresholding and Weight Decomposition

Abstract: Hierarchical convolutional neural networks are a well-known robust image-recognition model. In order to apply this model to robot vision or various intelligent real-time vision systems, its VLSI implementation is essential. This paper proposes a new algorithm for reducing multiply-and-accumulation operation by thresholding in a projection field and by performing weight decomposition in a 2-D neuron array. We also propose a VLSI architecture based on the proposed algorithm, and estimate its operation performance.

Neural networks with on-the-fly confidence-estimation

Abstract: Artificial neural networks are an established solution for classification. Many neural networks consist of perceptrons (e.g. multilayer-perceptron networks). In this paper, we propose a concept to extract reliability information of a perceptron-based neural networks decision. Therefore, we extend the conventional perceptron and suggest a simple training strategy.

A robust flight coupon recognition system based on multi neural network

Abstract: It is essential to recognize the flight coupon information rapidly and correctly in the modernization and information construction of civil aviation. Flight coupons are usually with complex background and multifont. In order to achieve a high recognition rate and reliability, we have designed our recognition system based on multi neural network. Both the back propagation neural network and the convolutional neural network are used in our multi neural network model, and this model can successfully discriminate the classes with similar features. As a real-time recognition system, we have a high character recognition speed. In this paper, we present the solution and the experimental result of the system.