Machine vision system

A machine vision system, comprising:

• a receiver configured to receive training data including image data indicative of an object and an associated known classification of the object;
• and,
• a processor provided with an initial neural network algorithm having an input layer provided with the received image data, at least one hidden layer and an output layer, the at least one hidden layer including a plurality of neurons each having an associated activation function, each connection between two of the neurons having an associated synaptic weight, and the processor being configured to determine appropriate values of the synaptic weights to ensure that evaluating the initial neural network algorithm with the received image data produces the associated known classification,
• memory in electronic communication with the processor;
• instructions stored in the memory, the instructions being executable to implement a method comprising:
• determining a single ordinary differential equation representing the entirety of a neural network describing the initial neural network algorithm based on the activation functions of the plurality of neurons, wherein in the single ordinary differential equation an output of a differential equation for a particular layer of the neural network is an input to a differential equation of a next layer of the neural network;
• determining a solution to the single ordinary differential equation in the form of a series expansion;
• converting the series expansion to a finite series expansion by limiting the number of terms in the series expansion to a finite number;
• determining an output classification in dependence on the finite series expansion;
• and,
• adjusting the values of the synaptic weights so as to reduce error between the determined output classification and the associated known classification of the object.

A machine vision control system for configuring a manufacturing process component/system, the machine vision control system comprising:

• a camera configured to receive image data indicative of an object to be classified;
• and,
• a processor provided with an initial neural network algorithm having an input layer provided with the received image data, at least one hidden layer and an output layer, the at least one hidden layer including a plurality of neurons each having an associated activation function, the processor being configured to provide an output classification to the output layer in dependence on the received image data;
• and,
• a controller configured to send a control signal to the manufacturing process component/system in dependence on the output classification so as to configure a setup of the manufacturing process component/system,
• memory in electronic communication with the processor;
• instructions stored in the memory, the instructions being executable to implement a method comprising:
• determining a single ordinary differential equation representing the entirety of a neural network describing the initial neural network algorithm based on the activation functions of the plurality of neurons, wherein in the single ordinary differential equation an output of a differential equation for a particular layer of the neural network is an input to a differential equation of a next layer of the neural network;
• determining a solution to the single ordinary differential equation in the form of a series expansion;
• converting the series expansion to a finite series expansion by limiting the number of terms in the series expansion to a finite number;
• and,
• determining the output classification in dependence on the finite series expansion.

A production line arrangement for inspecting a production line component to identify and locate damage, foreign objects or incorrectly fitted parts of the production line component, the production line arrangement comprising:

• an optical system including at least one camera configured to capture one or more sets of image data relating to the production line component;
• a machine vision control system comprising a receiver configured to receive image data indicative of an object to be classified, and a processor provided with an initial neural network algorithm having an input layer provided with the received image data, at least one hidden layer and an output layer, the at least one hidden layer including a plurality of neurons each having an associated activation function, the processor being configured to provide an output classification to the output layer in dependence on the received image data;
• memory in electronic communication with the processor;
• instructions stored in the memory;
• and,
• a manufacturing process component/system configured to adjust a set up thereof in dependence on the output classification,
• the instructions being executable to implement a method comprising:
• determining a single ordinary differential equation representing the entirety of a neural network describing the initial neural network algorithm based on the activation functions of the plurality of neurons, wherein in the single ordinary differential equation an output of a differential equation for a particular layer of the neural network is an input to a differential equation of a next layer of the neural network;
• determining a solution to the single ordinary differential equation in the form of a series expansion;
• converting the series expansion to a finite series expansion by limiting the number of terms in the series expansion to a finite number;
• and,
• determining the output classification in dependence on the finite series expansion.

A machine vision method, comprising:

• receiving image data indicative of an object to be classified;
• providing an initial neural network algorithm having an input layer provided with the received image data, at least one hidden layer and an output layer, the at least one hidden layer including a plurality of neurons each having an associated activation function;
• determining a single ordinary differential equation representing the entirety of a neural network describing the initial neural network algorithm based on the activation functions of the plurality of neurons and other associated neural network parameters, wherein in the single ordinary differential equation an output of a differential equation for a particular layer of the neural network is an input to a differential equation of a next layer of the neural network;
• determining a solution to the single ordinary differential equation in the form of a series expansion;
• converting the series expansion to a finite series expansion by limiting the number of terms in the series expansion to a finite number;
• and,
• providing an output classification to the output layer in dependence on the received image data, wherein the output classification is determined in dependence on the finite series expansion.