WO1998012518A2

WO1998012518A2 - Spatial photometric neural network

Info

Publication number: WO1998012518A2
Application number: PCT/IB1997/001223
Authority: WO
Inventors: John M. Suhan
Original assignee: Wea Manufacturing, Inc.
Priority date: 1996-09-05
Filing date: 1997-09-05
Publication date: 1998-03-26
Also published as: AU4316297A; WO1998012518A3

Abstract

A neural network algorithm permits the characterization of spatially distributed light reflected from the surface of a compact disc (1) and provides a means for processing such location-dependent outputs of photosite arrays. The outputs of a photosite array (11), such as a ring-wedge photodetector, are sent to the inputs of the neural network (13), where the designed transfer function yield output values that are characteristic of the spatial distribution of light upon the photosite array (11).

Description

SPATIAL PHOTOMETRIC NEURAL NETWORK

Field Of The Invention This invention relates to the detection of defects in the recording of optical information on a compact disc. In particular, the invention further relates to such detection by the classification of optical diffraction patterns by the use of a neural network.

Background Of The Invention Compact discs are data media having a metal layer formed by metal deposition over a transparent plastic substrate having pits of varying length. The data is encoded by the dimensions of the pits and can be read by a laser that passes through the transparent surface of the disc to reach the metal layer. On the reverse side of the metal layer additional material is deposited to build up the thickness of the disc. This reverse side need not be transparent and typically a label is printed there. The label often contains attractive art work and quality control procedures are necessary to ensure uniformity of the final product during mass production processes.

The image on the surface of the compact disc may be considered to be characterized as a spatial arrangement of photosite arrays. A great deal of trial and error techniques employing manual inspection an manipulation of the disc was required to determine the implications of a given optical pattern. For example, where the testing of parts for visual defects were made, scans would be taken for both the defective and the nondefective parts. The optical system responsible for the scans would image the surface of each part in a manner which projects the image of each surface across a photosite array.

Numerical values could be generated via a computer corresponding to each photsite of the array. The manual manipulation came at this point in the process, where the generation of computer code was required to make classifications, based on the experimental observations. With so many combinations of defect geometries, a very large amount of classifications were required such that all possibilities would be covered. Obviously, covering such a wide base with hard computer code went well beyond the capabilities of simple programming to the point where concessions were required to afford programmability.

The end result of these concessions was that many classification categories were ignored, causing an incomplete classification. Brief Description Of The Invention A Spatial Photometric Neural Network employs an automated scanning of a diffraction pattern formed from a projection of the image of the surface of a disc. The diffraction pattern in effect transforms the visual data into the Fourier transform of the original image in the plane of a focusing lens. A neural network is trained to compare the dominant Fourier coefficients of the transform of the image to a predetermined set and therefore to determine whether the image passes quality control.

Brief Description Of The Drawings Figure 1 is a schematic representation of the scanning apparatus of the present invention. Figure 2 shows the levels of nodes of the back propagation neural network.

Detailed Description Of A Preferred Embodiment

As shown in Figure 1, a disc to be tested 1_^ is illuminated by coherent light, such as emanates from a narrow point source of light 3_, or a laser device. The light is reflected onto a diffraction grating 5_ and focused by a lens 1_ having a focal plane j). The intensity distribution in the focal plane contains a group of primary, secondary, tertiary, etc. maxima which are detected by a photometer 11. This spectrum is representative of the image on the compact disc 1_ and represents a significant compression of the amount of information that must be analyzed to determine whether the image on the disc accurately reproduces the intended image in terms of its geometric configuration. The relationship between the original image and the intensity pattern at the photometer is that of an image to its Fourier transform.

The photometer is connected to a computer 13, which runs a neural network as described below. The purpose of the neural network is to accept training to learn the details of the original image and then to make the comparison between the image of each disc and the original image. A back-propagation neural network is preferred for this application due to its ability to readily adapt to the training process. As shown in Fig. 2, the neural network is a collection of logical nodes 12_ arranged in layers 1_3, 1_5, and 1/7 with the nodes in one layer connected to the nodes in many nodes in other layers. Layer 1/7 contains the input nodes, layer 1/3 the output nodes, and layer 1_5 the so-called hidden nodes. Each node processes the input it receives through these connections. The strengths of the connections changes in response to the strengths of the inputs and the transfer function used by the node. The transfer function mathematically expresses the relation between input and output. A neural network is defined by how its nodes are created, how the nodes process the information that they receive and how the connection strengths are modified. The preferred neural network of the present invention is a back-propagation feed-forward network. In this network data flows only in one direction from layer to layer. This is contrasted with feedback and recurrent networks in which the nodes are connected such that a later layer may provide information back to an earlier layer. The network of the preferred embodiment is a trained network. The training of the network is a procedure consisting of providing the network with typical expected inputs at an input layer and the desired outputs at the output layer. The nodes are then adjusted so that repetitions of these inputs will produce the desired outputs. The network is then "trained" in a supervised learning procedure termed Hebbian learning to provide similar outputs for similar inputs. Initially the network produces erroneous answers and an error is calculated. The error is used to adjust the weights in the network to approximate the correct response.

The procedure begins with the teaching process of what conditions constitute defects and what conditions constitute non-defects. Once this has been determined, scans to the photsite array must occur. A computer is used in the process of determining the optical intensity values at each photsite. The value as well as its array coordinates must be recorded for teaching the neural network.

The back-propagation neural network paradigm is preferred for this application due to its ability to readily adapt to the training process. The training process takes place by collecting several "sample" patterns (good and bad) , taking their respective spatial photometric data, and presenting this data to the inputs of the back-propagation neural network. An expected output (good or bad) for each pattern must be revealed to the output nodes of the neural network, as well.

Training of the network would occur under the typical provisions and requirements of the back-propagation neural network paradigm.

With training complete, testing can occur by presenting new photsite array values and addresses to the input nodes of the neural network. An output value will be generated by the neural network as a result of the presented input values. This output value is the quality status of the surface presented to the photsite array. The neural network can be automated by interfacing its computer in-line with the appropriate surface fabrication equipment. Defective parts can be screened directly on the production line.

The described Spatial Photometric Neural Network provides an efficient means for classifying optical diffraction patterns without the need for manual classification, sorting, and programming. The self-teaching system provides an accurate means of grouping defect classes. Such a neural network can be part of an in-line automated system in a man- ufacturing environment.

While there have been shown and described and pointed out the fundamental novel features of the invention as applied to preferred embodiments thereof, it will be understood that various omissions and substitutions and changes in the form and details of the device illustrated and in its operation may be made by those skilled in the art without departing from the spirit of the invention, which is exemplified in the following claims.

Claims

I claim:

1. An automated system for performing quality control for designs on compact discs comprising

(a) reflecting coherent light from the surface of a compact disc and directing the light through a diffraction grating and lens system to a photodetector,

(b) forming a diffraction pattern having primary and secondary diffraction maxima,

(c) inputting said points of values of diffraction maxima, to the input layer of a back-propagation neura1 network,

(d) training said neural network to reproduce values representative of an acceptable image pattern on said compact disc, (e) analyzing sample discs by repeating steps

(a) - (c) with said sample discs.

2. The automated system for performing quality control for design son compact discs of claim 1, wherein said lens has a focal plane and said photometer receives light from the focal plane of said lens.