Research On Infrared Image Processing Of Toxic Gases Based On Deep Learning Algorithm

Industrial toxic gases emission or leakage such as NH3,C12,SF6,cause lots of irreversible damages to human’s life and property.Therefore,it has great significance and value to recognize the type and concentration of the leakage gas correctly and efficiently.Infrared image processing becomes one of the most efficient technologies,due to it’s long distance,large range,high efficiency and dynamic visualization.In this paper,we process gas infrared image based on deep learning,it avoid image information losses during the general machine learning algorithms such as feature extraction,image enhancement,pseudo-color handling.Our algorithm achieves more than 98%accuracy rate,more than general machine learning method hog-svm algorithm,which can reach 92%accuracy.The main innovative work of this paper as follows:(1)Propose Non-local means(NLmeans)algorithm to denoise infrared image,then linear-transform 14bit raw infrared image to 8bit image,which can be displayed on computer screen.Compare NLmeans to the wavelet denoising algorithm.Analysis the influences of different search region radii,different neighbour region radii and smoothing coefficient parameters.Draw some conclusions as followed:1.NLmeans algorithm does better than wavelet in infrared image denoising;2.the bigger search region radius or different neighbour region radius or smooth coefficient parameter,the better denoising efficiency we get,but the image’s edge would be obscurer.(2)Combine No-reference Image Quality Assessment(NR-IQA)and deep learning algorithm.Train convolutional neural network by LIVE data set.Utilize the trained model to evaluate toxic gas’s quality.The performance of NR-IQA algorithm is evaluated by calculating the linear correlation coefficient(LCC)and the Spielman rank correlation coefficient(SROCC)of the regression value and the Difference Mean Opinion Score(DMOS).Compare the LCC and SROCC of CNN to other general NR-IQA algorithm.Draw some conclusions as followed:1.CNN does better than CORNIA or BRIQUSE in NR-IQA;2.NLmeans algorithm achieves the lower regression value than wavelet,which means that NLmeans algorithm does better than wavelet in image denoising.(3)Combine CNN to infrared image classification of toxic gases.Machine learning in classification domain contain feature extraction and bp neural network or SVM,which always cause the lack of image information.CNN extract more detail features by its deeper network construction which can avoid the lack of image information.This paper analysis that hog-svm algorithm reaches 92%accuracy in classification and our algorithm achieves more than 98%accuracy.We also analysis the effects of different dropout ratios,different learning rates and different optimizer algorithms on classification results.This paper achieves more than 98%accuracy rate.