Digitization Scheme Research Articles

Signal processing and training by a neural network for phoneme recognition

An automated speech recognition system converts a speech signal into a compact, coded representation that correlates to a speech phoneme set. A number of different neural network pattern matching schemes may be used to perform the necessary speech coding. An integrated user interface guides a user unfamiliar with the details of speech recognition or neural networks to quickly develop and test a neural network for phoneme recognition. To train the neural network, digitized voice data containing known phonemes that the user wants the neural network to ultimately recognize are processed by the integrated user interface. The digitized speech is segmented into phonemes with each segment being labelled with a corresponding phoneme code. Based on a user selected transformation method and transformation parameters, each segment is transformed into a series of multiple dimension vectors representative of the speech characteristics of that segment. These vectors are iteratively presented to a neural network to train/adapt that neural network to consistently distinguish and recognize these vectors and assign an appropriate phoneme code to each vector. Simultaneous display of the digitized speech, segments, vector sets, and a representation of the trained neural network assist the user in visually confirming the acceptability of the phoneme training set. A user may also selectively audibly confirm the acceptability of the digitization scheme, the segments, and the transform vectors so that satisfactory training data are presented to the neural network. If the user finds a particular step or parameter produces an unacceptable result, the user may modify one or more of the parameters and verify whether the modification effected an improvement in performance. The trained neural network is also automatically tested by presenting a test speech signal to the integrated user interface and observing both audibly and visually automatic segmentation of the speech, transformation into multidimensional vectors, and the resulting neural network assigned phoneme codes. A method of decoding such phoneme codes using the neural network is also disclosed.

Automation of the axisymmetric drop shape analysis-diameter for contact angle measurements

Axisymmetric drop shape analysis-diameter (ADSA-D) was developed to measure contact angles on non-ideal and/or hydrophilic surfaces (e.g. biological surfaces). ADSA-D determines the contact angle by solving the Laplace equation of capillarity numerically, and using as input the liquid surface tension, the drop volume and diameter of the drop as viewed from above. The diameter is the main experimental parameter to be determined as the liquid surface tension is known and the volume of the sessile drop can be obtained precisely by means of the micrometer syringe used to form the drop. The edge detection method to determine the diameter, which consisted of manually extracting coordinates from a digital image was found to be time consuming and somewhat subjective. Therefore, an automated image processing module was developed to replace the manual scheme. This automated module uses nonlinear filters to process the images and uses a region growing scheme to determine the total area of the drop as viewed from above. This area is then used to calculate the equivalent diameter of the idealized contact or equatorial circle for the drop. The robust design of the automated module enables it to process accurately many different types of drop images. Two very different examples of water drops on bacteria and carboxylic acid terminated self assembled surfaces are used to demonstrate the capabilities of the new image processing scheme. Also, using these two examples, it is shown through contact angle measurements that the results of this new ADSA-D version are more accurate than those obtained through the manual digitization scheme previously used.

Method and apparatus for interfacing and training a neural network for phoneme recognition

An automated speech recognition system converts a speech signal into a compact, coded representation that correlates to a speech phoneme set. A number of different neural network pattern matching schemes may be used to perform the necessary speech coding. An integrated user interface guides a user unfamiliar with the details of speech recognition or neural networks to quickly develop and test a neural network for phoneme recognition. To train the neural network, digitized voice data containing known phonemes that the user wants the neural network to ultimately recognize are processed by the integrated user interface. The digitized speech is segmented into phonemes with each segment being labelled with a corresponding phoneme code. Based on a user selected transformation method and transformation parameters, each segment is transformed into a series of multiple dimension vectors representative of the speech characteristics of that segment. These vectors are iteratively presented to a neural network to train/adapt that neural network to consistently distinguish and recognize these vectors and assign an appropriate phoneme code to each vector. Simultaneous display of the digitized speech, segments, vector sets, and a representation of the trained neural network assist the user in visually confirming the acceptability of the phoneme training set. A user may also selectively audibly confirm the acceptability of the digitization scheme, the segments, and the transform vectors so that satisfactory training data are presented to the neural network. If the user finds a particular step or parameter produces an unacceptable result, the user may modify one or more of the parameters and verify whether the modification effected an improvement in performance. The trained neural network is also automatically tested by presenting a test speech signal to the integrated user interface and observing both audibly and visually automatic segmentation of the speech, transformation into multidimensional vectors, and the resulting neural network assigned phoneme codes. A method of decoding such phoneme codes using the neural network is also disclosed.

Method and apparatus for developing a neural network for phoneme recognition

An automated speech recognition system converts a speech signal into a compact, coded representation that correlates to a speech phoneme set. A number of different neural network pattern matching schemes may be used to perform the necessary speech coding. An integrated user interface guides a user unfamiliar with the details of speech recognition or neural networks to quickly develop and test a neural network for phoneme recognition. To train the neural network, digitized voice data containing known phonemes that the user wants the neural network to ultimately recognize are processed by the integrated user interface. The digitized speech is segmented into phonemes with each segment being labelled with a corresponding phoneme code. Based on a user selected transformation method and transformation parameters, each segment is transformed into a series of multiple dimension vectors representative of the speech characteristics of that segment. These vectors are iteratively presented to a neural network to train/adapt that neural network to consistently distinguish and recognize these vectors and assign an appropriate phoneme code to each vector. Simultaneous display of the digitized speech, segments, vector sets, and a representation of the trained neural network assist the user in visually confirming the acceptability of the phoneme training set. A user may also selectively audibly confirm the acceptability of the digitization scheme, the segments, and the transform vectors so that satisfactory training data are presented to the neural network. If the user finds a particular step or parameter produces an unacceptable result, the user may modify one or more of the parameters and verify whether the modification effected an improvement in performance. The trained neural network is also automatically tested by presenting a test speech signal to the integrated user interface and observing both audibly and visually automatic segmentation of the speech, transformation into multidimensional vectors, and the resulting neural network assigned phoneme codes. A method of decoding such phoneme codes using the neural network is also disclosed.

Digital circles

Digital circles, which are the digital images of real circles, are introduced. A necessary and sufficient condition for a given finite set of lattice points (or cells) to be the digital image of a circle providing that the center of the circle is given is investigated first. Since the definition of digital circles is given in terms of preimages, the condition depends on the digitization scheme used as well as on the geometrical properties of the set. Thus, three different digitization schemes which have been used for arcs will be used here. Then an algorithm is given to detect the center in a finite number of steps. By combining these results, algorithms that determine whether or not a given finite set of lattice points (or cells) is a digital image of a circle under three different digitization schemes are presented.

More about areas and perimeters of quantized objects

A controversy between two common approaches to area and perimeter measurement of discrete blobs is resolved in this paper. The first approach takes the chain of 8-connected border pixels as the object boundary (thus using Pick's theorem to exclude a contribution of a part of these to the value of the area), whereas the second one considers the boundary of the object to be a line separating object-border pixels from background ones (thus counting all pixels of the object as its area). It is shown that the validity of these approaches (and corresponding measurement methods) is related to the digitization scheme assumed: the first approach is valid for a boundary-line digitization (or edge-detection) scheme, and the second one is better for a point-sampling digitization (or region-extraction or thresholding) scheme.

Region digitization and boundary estimation

Digital complexes most often arise in practice as binary images, digitized by computer vision systems from real objects. In this paper, schemes for region digitization and boundary estimation are discussed in relation to the actual physical process of binary image acquisition. The region digitization scheme motivated naturally by such a discussion is also appropriate for use in theoretical studies of digital complexes. The same discussion suggests a method for obtaining an estimate of the boundary of an object from its binary image.

Regional Patterns of Mixing Depth and Stability: Sodar Network Measurements for Input to Air Quality Models

Abstract A network of 13 sodars (acoustic radars) was operated in the San Francisco Bay Area for 2.5 months of the 1976 smog season. The goal was to produce a data base on time-dependent mixing depth and stability patterns for input to air quality models. The large set (∼1000 site-days) of sodar facsimile records was filmed to provide a more compact, convenient, and accessible data product. A manual digitization scheme was devised and used to convert the continuous sodar data to hourly values describing mixing depth and a near-surface stability indicator. These values can easily be input to computerized air quality models and can readily be understood by a user with no sodas experience. Numerous sodar-inferred mixing depths were compared to those inferred from 1) simultaneous measurements of temperature and humidity profiles and 2) lidar-measured haze and cloud layering. These and previous tests show good overall agreement, demonstrating that in the San Francisco Bay Area, sodar measurements compare very ...

Area and perimeter measurement of blobs in discrete binary pictures

In picture processing one is often confronted with the problem of deriving some geometrical parameters (e.g., area, perimeter) of continuous figures (blobs, curves) from their digitizations. This paper concerns some particular aspects of this problem, namely, the construction of practical, simple, and computationally fast methods for measuring accurately the area and perimeter of blobs given as digital binary pictures. The analysis of some properties of digital pictures leads to two such methods. The first one uses only local, position-invariant parallel picture processing operations. The second one uses explicit polygonal approximations of the picture contour, with controlled degree of approximation, and can work on more noisy pictures. The methods were practically tested in a working program for shape factor measurement of fiber sections. In addition some useful properties of Freeman's digitization scheme are investigated and proved.