Pulse phase infrared thermography combines the advantages of both pulse infrared thermography and modulated infrared thermography. In this project, the use of a neural network to perform non destructive evaluation of materials using the phase spectrum obtained by pulse phase infrared thermography will be studied. This text summarizes the avenues that will be explored to achieve the goals of this project.

Within the context of the Computer Vision and Systems Laboratory research, we will study the use of a neural network to perform non destructive evaluation of materials using the phase spectrum obtained by pulse phase infrared thermography. This recently proposed method brings together the advantages of both conventional methods, pulse infrared thermography and modulated infrared thermography. A preliminary study on the use of neural networks by Yann Largouet [1] shows that neural network processing using the phase obtained by pulse phase infrared thermography could have many important practical applications. This project is only beginning, so we will present here a summary of its essential points.

The theoretical study done by Y. Largouet shows that problems of noise emerge. The phase spectrum is more sensitive to the noise than the sampled temperature values. Indeed, the first experimental phase images observed were significantly distant from the theoretical profile. To obtain good results, the effect of noise has to be minimized.

First, the use of an unsupervised learning Kohonen neural network is considered because this type of neural network exhibits a better noise immunity than the perceptron network used by Y. Largouet in his study. Second, Y. Largouet observed that the type of material influences the quality of the results. The three layer perceptron used in his study did not behave properly to materials with high heat diffusion. This type of material is more sensitive to noise and requires a faster sampling frequency. Hence the frequency needed for the sampled images will be studied : we observed theoretically that doubling the sampling frequency results in a decrease of noise effects and an improvement of the quality of results for the same material.

On the other hand, we think that the use of the phase along with the first values of amplitude as input to the neural network could help in obtaining better results without increasing significantly the number of neurons. The first values of amplitude possess a large part of the total energy, and should give additional information on the observed sample.

This project will confirm experimentally the results of the theoretical study made by Y. Largouet. The goal of this project is to realize a system with a good immunity to noise. This system will be able to extract quantitative information about the possible flaws in a sample.

a) real image        b) amplitude image
c) maximum thermal contrast      d)	phase image
Figure 2. Images from an airplane part.

Reference:
[1] Yann Largouet, Rapport de stage, LVSN, sept. 1997.