Characteristics Extraction Of Metallic Corrosion Morphology For Corrosion Diagnosis

Corrosion images and corrosion data are very important and fundamental to diagnose corrosion type, evaluate corrosion degree and study corrosion mechanism in corrosion engineering. Manual identification and detection of these images is a tedious work. It is also easy to be affected by subjective factors. Thus, quantitatively analyzing corrosion morphology, extracting and transforming the corrosion information from numerous corrosion images into recognizable information by computers to obtain the useful knowledge is one of the most important subjects needed to be addressed in corrosion engineering.In this paper, extraction of corrosion characteristics from various corrosion morphology images of 20# carbon steel, 5454 Al alloy, 304 stainless steel and industrial pure Al have been carried out employing several image processing techniques such as gray statistic, wavelet transform and binary technique. A new method for corrosion type identification and corrosion rate prediction is developed based on the corrosion characteristics extraction combining with Back Propagation (BP) neural networks. Fractal theory is also employed to describe the irregular morphologies and the complex pits distribution of corroded surface, which gives a new approach for diagnosing corrosion type and corrosion degree.Firstly, corrosion image pre-processing and segmentation are investigated. The corroded surface morphology images of 5454 Al alloy in the industrial circulation-water environment are digitized to gray matrixes. The contrast between corroded region and material matrix is enhanced by histogram stretching and the corrosion edges are detected by Sobel and Prewitt methods.By selecting pitting images of 304 stainless steel which obtained by immerse method and electrochemical experiments both exposed in FeCl3 solutions as tested morphology images, corrosion features contained in these images are extracted using three typical methods including gray statistic, wavelet transform and binary characteristics. Pitting area and pits number have been obtained accurately using binary corrosion image processing based on 8-connected algorithm. Values of several statistical feature parameters such as the mean gray, the standard deviation, the energy and the entropy of image grays, are calculated based on the gray matrixes ofthe interest images. These parameters are available to describe the undulation corroded morphology and information distribution over the corroded surface. The original corrosion image can be decomposed as low frequency sub-images and high frequency sub-images by wavelet transform. It is helpful not only for segmenting the corroded region from the material base but also for decreasing calculation of the gray matrix.Secondly, anisotropy energy and energies deduced from wavelet sub-images of 304 stainless steel and carbon steel corroded surface images in FeCl3 solution are considered as input factors in a BP neural network which is then applied to identify pitting and general corrosion. The diagnosis results agree well with the experimental corrosion types. A 2-5-1 BP neural network model is developed for the prediction of pitting corrosion rate of 304 stainless steel by using the obtained data of the corrosion ratio and pits density obtained by chemical immerse experiments. There is a good agreement between the predicted results and the experimental data of pitting corrosion rate. The max relative error of prediction is 8.81%. In the same way, A 6-8-2 BP neural network model is developed for the prediction of pitting corrosion rate and pit deepness of 304 stainless steel by using the obtained data such as concentration and temperature of FeCl3 solution, and the characteristic factors of corrosion image including corrosion ratio, pits density and fractal dimension of pits distribution obtained by electrochemical testing. There is also a good agreement between the simulating results and the experimental data of pitting corrosion rate and pit deepness. The max relative errors of simulation are 6.69% and 4.62%, respectively.Finally, gray a...