Average Feature Values Research Articles

The integration of discriminative and maximum likelihood distance measures for a vowel articulation training aid

Improvements to the vowel articulation training aid described in a previous paper [Zimmer, Zahorian, and Auberg, J. Acoust. Soc. Am. 101, 3199(A) (1997)] have been made. The system uses a standard Windows 95/NT compatible sound card on a multimedia PC to provide continuous feedback about articulation for ten American English monopthong vowels in two modes: an F1/F2 style ‘‘ellipse’’ display and a vowel bargraph display. Neural network discriminative classifiers are used to produce the display outputs based on 12 features. Though testing showed that the system provided useful output for most voiced speech sounds, the network would classify some out-of-category sounds (e.g., nonvowel sounds, ambient noise) as belonging to one of the ten vowel categories. Experimental tests indicate that the inclusion of a generalized Euclidean distance measure to compare the feature values of an utterance with the average feature values for the vowel category specified by the neural network output help to greatly reduce the number of out-of-category sounds improperly classified as ‘‘correct’’ responses. The paper will describe the processing in more detail and summarize experimental results from vowel classification tests. [Funded by NSF, Grant No. NSF-BES-9411607.]

Estimation of Fusarium Scab in Wheat Using Machine Vision and a Neural Network

ABSTRACTA neural network was used to relate color and texture features of wheat samples to damage caused by Fusarium scab infection. A total of 55 color and texture features were extracted from images captured by a machine vision system. Random errors were reduced by using average values of features from multiple images of individual samples. A four‐layer backpropagation neural network was used. The percentage of visual scabby kernels (%VSK) estimated by the trained network followed the actual percentage with a correlation coefficient of 0.97; maximum and mean absolute errors were 5.14 and 1.93%, respectively. A comparison between the results by the machine vision‐neural network technique and the human expert panel led to the conclusion that the machine vision‐neural network technique produced more accurate determination of %VSK than the human expert panel.

Effects of Element Density on Segmentation by Luminance, Colour, and Motion

A target composed of random dots can be easily segmented from a background of random dots when each is defined by a different, broad distribution of dot speeds, colours, or brightnesses, even if these distributions overlap extensively. This shows that segmentation can be based on global properties of the target and background. If average feature value is the relevant global property, segmentation threshold should be a monotonically decreasing function of dot density. Observers viewed a 6.5 deg × 6.5 deg square array of between 512 and 36 864 randomly positioned dots (dot size 4.7 or 1.6 min arc) in which the target was a vertical band of dots defined by a different distribution of luminance, colour, or motion from the background. The task was to indicate whether the target appeared on the left or on the right of midline. The stimulus was displayed for 67 ms, followed after 233 ms by a 67 ms mask. We found that motion segmentation thresholds decrease monotonically as a function of density. For luminance and colour, segmentation thresholds initially decrease toward a local minimum, then increase to a local maximum. Thereafter, thresholds decrease monotonically as density increases further. These results suggest that the integrative segmentation process for motion is qualitatively different from those for luminance and colour. The local extrema for luminance and colour suggests that segmentation in these domains is based on forming clusters by linking together dots with similar feature properties. Only when interdot distance is sufficiently small are all dots within an area included in the cluster. Behaviour is then determined by difference in average feature value or target — background contrast.