초록 ▼

Artificial Intelligence (AI) techniques are used for processing data having multiple possible values, for example, gray-scale values of pixels of an image. This is useful in scanning of digitally-coded image data to detect anomalies, for example defects in some of a series of manufactured products.

Artificial Intelligence (AI) techniques are used for processing data having multiple possible values, for example, gray-scale values of pixels of an image. This is useful in scanning of digitally-coded image data to detect anomalies, for example defects in some of a series of manufactured products. The preferred processing apparatus has either one processing channel or multiple parallel processing channels. The apparatus samples or scans incoming signals to produce a matrix of scan values, selects subsets of the scan values to define samples, extracts at least one feature, such as mean image brightness (MM), activity (MA), gradient absolute value (MG) and/or gradient direction (MD) from each sample, uses the at least one feature as an address to retrieve, from a table, a corresponding probability-of-occurrence value, derives, from the probability-of-occurrence values of a plurality of the samples, a decision value, and compares the decision value with a predetermined threshold value. The probability-of-occurrence values are stored in the table during a “learning”procedure.

대표청구항 ▼

Method for classification of signals, which signals each have more than two possible amplitude values, each sequence of signals collectively defining image background information and a feature, using data processing means having at least one processing channel (K1-K32), comprising the steps of: carr

Method for classification of signals, which signals each have more than two possible amplitude values, each sequence of signals collectively defining image background information and a feature, using data processing means having at least one processing channel (K1-K32), comprising the steps of: carrying out a learning process which generates and stores, at respective memory addresses in a memory of said data processing means, a table consisting of a plurality of probability-of-occurrence values of respective features, each occurrence of a particular feature during said learning process causing incrementing of the stored probability value of the feature; sampling incoming signals to produce a matrix of scan values; selecting subsets of scan values to define samples; extracting at least one feature from each sample; using said at least one feature as an address to retrieve, from said table, a corresponding one value from among said plurality of probability-of-occurrence values; deriving from said one probability-of-occurrence value for each one of a plurality of said features respectively corresponding to a plurality of said samples, a decision value; comparing said decision value with a predetermined threshold value for classifying the signals; and subjecting said signals to a window function prior to said step of extracting features from said samples, said window function including the step of passing only ones of said samples within a predetermined range of values for further processing by said step of extracting.