Design And Optimization Of Intelligent Light Source For Machine Vision Application

Light source is the principal factor of machine vision's input data.It directly affects the quality of input data and 30% application effect.A proper light source(including color,illumination,lighting angles)enables the camera to obtain high quality images,so as to simplify the subsequent algorithm of image processing and enhance the accuracy and reliability of machine vision.But some problems focus on machine vision's light source still exit currently,for example,to improve image contrast,one type of light source needs different illumination colors according to await measuring characteristic of color.Different lighting conditions are required under different working illumination,and the influence of ambient light on vision is different between day and night,which may lead to instability of detecting system.Most light source keep lighting during the work of visual system and this may result in long working hours,heating severely and energy wasting.In practice,an industrial camera only need proper light at a specific time.To solve these problems,this paper studies around three aspects: lighting system,PWM digital control and neural network.It exploit the control algorithm of PWM,and add recognition algorithm of neural network.Each part has experimental verification.This paper introduces and summarizes the common way of lighting method,and also put forward a kind of digital control light source based on PWM.The function relation between the three channels PWM digital numbers and light intensity is established.Meanwhile,to introduce the working principal of light source from four aspects: dimming method,illumination control,ATmega2560 singlechip as well as design of circuit.A mathematical model for the mixed light mode is established,and the MFC interface is developed based on Visual Studio.Finally,this mathematical model which is based on single channel,dual channel and three-channel has been tested.The experimental results show that the model can meet the majority of the mixed light model and realize the color and illumination of the light source.On this basis,the scenes of the light source is divided into three conditions: over light,over dark and normal.Based on the tensorflow platform,the neural network algorithm will be added to light source which make itself can judge the scene how it is,according to the obtained images.In the training of neural network model,the learning rate,initialization weights,number of neurons,and Droup Out rate will be optimized.In the same condition,the experiments between SGD and Adam will be experiment contrastively.The comparison parameters include the rate of accuracy,maximum,minimum,mean,and variance of the bias and weight of the final training model.The experimental results show that two kinds of optimizers can achieved an accuracy of 99%,but the weight and the bias values have different solutions.