Research On Temperature Measurement Method Based On Visible Light Image And Machine Learning

Accurate temperature measurement is of great significance in many situations,such as industrial production,scientific experiments,equipment failure detection and so on.The thermal radiation of low-temperature objects in the visible band is so weak that infrared thermometer is mostly used for temperature measurement.However,the infrared image has much lower pixels and much lower spatial resolution than the visible light image.At present,the function of fault location and temperature measurement can only be accomplished by infrared image and visible image fusion.In this study,aiming at the situation that the faults of power equipments are basically at normal temperature,we try to realize temperature measurement by using visible light image directly,and establish a prediction model between visible image characteristics and temperature through machine learning technology.The specific results obtained are as follows.A low temperature measurement method based on K-nearest neighbor algorithm and visible light image is proposed.Firstly the visible light image library under different temperatures is established.The RGB Gray Level Histogram of visible light image captured by digital camera is extracted as the color features.Then linear regression,decision tree,gradient boosting decision tree,support vector machine and K-nearest neighbors algorithm are used to build the temperature prediction model by learning from color features.Finally the prediction accuracy is determined by cross validation method.The results show that the K-nearest neighbors algorithm model has the highest accuracy.The maximum absolute error between the predicted temperature and the real temperature is 4.3?.The mean absolute error between the predicted temperature and the real temperature is 0.023?.The influence of object color and object material on the prediction model is studied.We get three visible light image libraries by applying yellow,blue and green stickers to the surface of object.The maximum absolute errors of K-nearest neighbors temperature prediction model are 5.6?,2.883? and 5.938? respectively.We get two visible light image libraries of aluminum and brass respectively.The maximum absolute errors of the K-nearest neighbors temperature prediction model are 7? and 5.32? respectively.The results indicates that the temperature prediction method in this paper is feasible.The method to overcome the influence of ambient light variations on the prediction model is studied.The gray-world method,perfect-reflection method and retinex theory are respectively applied to the visible light image under different ambient light conditions,to mitigate the impact of the light variations on the RGB Gray Level Histogram of the visible light image.After Retinex method is adopted,the prediction error of K-nearest neighbors temperature prediction model is significantly reduced.In this paper,we proposed a low temperature measurement method based on K-nearest neighbor algorithm and visible light image which has high accuracy,simple operation and low cost.This method is an effective means for the temperature measurement of low temperature objects.