Methods of inspection

Claim 2

An inspection method comprising steps of picking up an image of an object to be inspected by a visual sensor to output a video signal; digitizing said video signal to obtain digital image data; calculating feature quantities based on said digital image data; inputting said feature quantities into a neural network comprising an input neuron, which is an adaptive propagation and vector quantization neuron, a plurality of intermediate neurons, each uniquely corresponding to one of quantization ranges present in said input neuron and a plurality of output neurons the number of which is equal to that of outputs to be needed; quantizing said feature quantities based on predetermined quantization ranges in said input neuron, calculating coupling coefficients between said input neuron and each of said plurality of intermediate neurons based on said feature quantities quantized in accordance with said predetermined quantization ranges to output a first output signal determined by said coupling coefficients to said plurality of intermediate neurons, obtaining a second output signal in each intermediate neuron by multiplying said input first output signal with an intermediate coupling coefficient between said each intermediate neuron and one of said plurality of output neurons which has been determined by learning to output said second output signal to the corresponding output neuron, converting said second output signals input from respective intermediate neurons into judging data using a function of said second output signals as variables in each of said plurality of output neurons; and judging whether said object is nondefective or defective by comparing judging data output from said plurality of output neurons with each other wherein said intermediate coupling coefficients are set so that any object is judged to be defective before starting learning and are fitted so that if an object is nondefective then it is judged to be nondefective by performing learning with use of nondefective objects.

Claim 3

The inspection method according to claim 1 wherein said quantization ranges locating on or near a border expected to be between an area of feature quantities to be judged defective and that to be judged nondefective are narrowed and other quantization ranges are widened.

Claim 4

The inspection method according to claim 1 further comprising a step of converting calculated feature quantities with scale conversion tables.

Claim 5

An inspection method comprising steps of picking up an image of an object to be inspected by a visual sensor to output a visual signal; digitizing said video signal to obtain digital image data; calculating feature quantities based on said digital image data; inputting said feature quantities into a neural network comprising an input neuron, which is an adaptive propagation and vector quantization neuron, a plurality of intermediate neurons, each uniquely corresponding to one of quantization ranges present in said input neuron and a plurality of output neurons including a first output neuron for outputting a maximum value among said plurality of output neurons if an object is said plurality of output neurons which has been determined by learning to output said second output signal to the corresponding output neuron; converting said second output signals input from respective intermediate neurons into judging data using a function of said second output signals as variables in each of said plurality of output neurons; and judging whether said object is nondefective or defective by comparing judging data output from said plurality of output neurons with each other; wherein said intermediate coupling coefficients are set initially so that said first output neuron outputs a maximum value among said plurality of output neurons for input feature quantities of an object within a range expected defective and a second output neuron for outputting a maximum value among values output from said plurality of output neurons if an object is nondefective; quantizing said feature quantities based on said present quantization ranges in said input neuron; calculating coupling coefficients between said input neuron and each of said plurality of intermediate neurons based on said feature quantities quantized to output a first output signal determined by said coupling coefficients to said plurality of intermediate neurons; obtaining a second output signal in each intermediate neuron by multiplying said input first output signal with an intermediate coupling coefficient between said each intermediate neuron and one of to be defective and said second output neuron outputs a maximum value among said plurality of output neurons for input feature quantities of an object within a range expected to be nondefective.

Claim 6

An inspection method comprising steps of picking up an image of an object to be inspected by a visual sensor to output a visual signal; digitizing said video signal to obtain digital image data; calculating feature quantities based on said digital image data; inputting said feature quantities into a neural network comprising an input neuron, which is an adaptive propagation and vector quantization neuron, a plurality of intermediate neurons, each uniquely corresponding to one of quantization ranges present in said input neuron and a plurality of output neurons including a first output neuron for outputting a maximum value among said plurality of output neurons if an object is said plurality of output neurons which has been determined by learning to output said second output signal to the corresponding output neuron; converting said second output signals input from respective intermediate neurons into judging data using a function of said second output signals as variables in each of said plurality of output neurons; and judging whether said object is nondefective or defective by comparing judging data output from said plurality of output neurons with each other; wherein said intermediate coupling coefficients are set initially so that said first output neuron outputs a maximum value among said plurality of output neurons for input feature quantities of an object within a range expected defective and a second output neuron for outputting a maximum value among values output from said plurality of output neurons if an object is nondefective; quantizing said feature quantities based on said present quantization ranges in said input neuron; calculating coupling coefficients between said input neuron and each of said plurality of intermediate neurons based on said feature quantities quantized to output a first output signal determined by said coupling coefficients to said plurality of intermediate neurons; obtaining a second output signal in each intermediate neuron by multiplying said input first output signal with an intermediate coupling coefficient between said each intermediate neuron and one of to be defective and said second output neuron outputs a maximum value among said plurality of output neurons for input feature quantities of an object within a range expected to be nondefective.

Claim 7

The inspection method according to claim 4 wherein said quantization ranges locating on or near a border expected to be between an area of feature quantities to be judged defective and that to be judged nondefective are narrowed and other quantization ranges are widened.

Claim 8

The inspection method according to claim 4 further comprising a step of converting calculated feature quantities with scale conversion tables.