SOME ISSUES IN THE AUTOMATIC DECODING OF RADIOGRAPHIC IMAGES

V. P. ABRAMOV, V. I. GORBUNOV

Recently, researchers’ interest in the development and creation of devices that replace the operator in decoding defectoscopic information has noticeably increased. The need for such devices is due to the following circumstances.

First, an automatic device will compare favorably with a human operator in the speed of processing and analyzing incoming information, which makes it possible to sharply increase inspection productivity.

Second, with the aid of automatic devices it is possible to achieve greater reliability in the assessment of quality than if this work were performed by an operator, since the machine deals with more precise quantitative relationships than a person does.

Third, if information on the causes of the occurrence of various disturbances in the internal integrity of the inspected products is incorporated into such devices, then, by closing the cycle from manufacture of the product to its inspection by feedback, it is possible to make the necessary changes in the technological process of manufacturing the product in order to exclude the appearance of dangerous defects, that is, to actively control production.

This problem appears especially urgent as applied to radiography, since this inspection method, being the most thoroughly studied, has become widely used in the practice of nondestructive testing of materials both at enterprises in this country and abroad. Existing precedents in the literature confirm the expediency of carrying out such developments. Thus, for example, in reports by the American firm “Lockheed” [1], whose work is among the first in the field of creating automatic devices for decoding images, it is stated that the device developed by the firm performs analysis and decoding of 30 m of X-ray film in 40 min, whereas a person performs the very same work in 7½ hours. For classification of defects, such features as size, location, and distance between defects are used. No data are given on the probability of correct classification of defects, their detection, or the sensitivity of the device; meanwhile, such characteristics are among the most important when comparing an automatic device with a person.

All of the foregoing was the reason why, at the Research Institute of Electron Introscopy, investigations are currently being carried out, the main purpose of which consists in developing an automatic machine that would, by means of

in decoding performance, was several times superior to a human being, while in sensitivity and correctness of the conclusion about quality it was at the level of an experienced human decoder.

The operating principle of the automatic device being developed is based on solving three tasks that are currently solved by a human operator when examining radiographic images, namely: the task of detecting defects recorded on X-ray film, the task of recognizing images of defects (determining the type and principal parameters), and the task of making a decision about the quality of the materials examined, taking into account the information from the solution of the first two tasks and also considering the future operating conditions of the product (operating mode, nature of the applied loads, etc.). A block diagram of the automatic device is shown in Fig. 1.

The results of inspection (in our case, X-ray films) enter the input of scanning device 1, whose main function is reduced to representing the specific physical characteristics of the object (features), that is, the optical density of radiographic images, in the form of a set of electrical signals. The operating principle of the scanning device and some results of its experimental investigations are considered in paper [2], included in the present collection. The video signal taken from the output of the scanning device, which reflects the optical density of the X-ray film at each point, is quantized with respect to the argument and with respect to the value of the function by quantization block 2.

Fig. 1. Block diagram of an automatic device for decoding radiographic images

The sampled values of the video-signal amplitudes obtained at intervals of time quantization enter detection block 3, which, in accordance with the selected quality criterion, optimally decides whether useful information is present in the received realization, that is, whether there are defects on the viewed X-ray film, or whether the received realization is noise, that is, there are no defects. The operating algorithm of the detection block is reduced to the following.

Initially, the mean value of the video-signal voltage is calculated for the entire radiographic image. For this purpose the discrete values of the video-signal amplitudes are summed, and the resulting sum \(\sum_{i=1}^{n} u_i\) is divided by the number of samples \(n\). Next, the relative value at each sample point is calculated, equal to

\[ \frac{U_i - U_{\mathrm{cp}}}{U_{\mathrm{cp}}} = \delta, \]

and already in the space of the quantity $\delta$ the information is separated into two classes: signal and interference. The choice of the quantity $\delta$ as the parameter by whose value the decision is made as to whether useful information is present (that is, whether there are defects in the image) or whether the image is a uniform background without regions of local changes in optical density makes it possible to achieve invariance with respect to the type of film and the development conditions. An error-free decision in this problem of detecting a useful signal cannot be obtained, since the probability-density distribution functions of the values of the feature $\delta$ for the signal and noise regions will intersect; that is, errors of the first and second kinds are possible. However, the well-developed methods of the theory of statistical decisions make it possible, using one or another quality criterion, to decide in an optimal manner: signal—accepted realization, or noise. The choice of the quality criterion for decision making is determined mainly by the available a priori information.

In developing the detection block in this problem, we used the Neyman–Pearson criterion, which makes it possible, for a given probability of false alarm, to minimize the probability of missing a signal from a defect.

Useful information from the output of the detection block enters recognition block 4, whose operating algorithm reduces to a detailed analysis and classification, into a specified number of classes, of the incoming information. First, the recognition block computes the specific physical characteristics of the objects (features), that is, the characteristics by which one class of objects differs from another. In choosing features, one should strive to ensure that, for a specified recognition probability, their number is minimal, for this determines the complexity of the decision-making device. In the present problem, the feature “contour shape” of the image of the detected defect was taken as the principal feature by which defects can be divided into the required number of classes (types), since it is known [3, 4] that certain types of defects have a characteristic shape. Quantitatively, this feature is expressed through various dimensionless ratios: “aspect ratio,” “ratio of the perimeter to the mean radius,” etc. Thus, for each class of defects there will be its own interval of values of the chosen feature. Then the falling of the computed value of the feature into one or another interval will determine the class of the detected defect. Second, the recognition block computes the necessary parameters of the detected defects: the linear size of the defect, planar dimensions, and the number of detected defects.

Thus, from the output of the recognition block comes information about the type of the detected defects, their parameters, and their number. This information is fed to the input of block 7, which, using the incoming information and taking additional conditions into account, issues a conclusion about quality, determining the degree to which the article may be used subsequently (the article is either rejected or deemed suitable for use under specified conditions).

In addition, such an automatic device can be supplemented with an information-display device of the cathode-ray-tube type 5, on whose screen images of the detected defects, cleaned of interference, and their necessary parameters will be displayed. Then, in those cases when the automatic device is uncertain in making a decision, this can be done by operator 6, viewing the CRT screen. In addition, the operator, using the CRT, can monitor the correctness of the operation of the blo-

blocks 4 and 7 and make the corresponding corrections to the algorithms of their operation, if this proves necessary; that is, the system will be trained in order to eliminate erroneous decisions.

In conclusion, we note that the development and creation of devices for the automatic analysis and recognition of radiographic images will make it possible either to exclude the operator entirely from the process of decoding the images, or to reduce his functions to monitoring the correctness of the system’s operation, which will significantly facilitate his work and increase the reliability and productivity of the radiographic inspection method.

REFERENCES

1. Computer reads weld X-rays, by Nicholas Srabo and Walter Eppler, La ckheed Missile Space CO, USA, Metalworking Production, october 13, 1965.

2. V. P. Abramov, V. I. Gorbunov, G. V. Nadokhovskii. Some results of a study of a device for reading information from radiographic images. Present collection.

3. S. T. Nazarov. Methods for quality control of welded joints. Mashinostroenie, Moscow, 1964.

4. Nondestructive testing. Handbook edited by R. McMaster, Energiya, Moscow—Leningrad, 1965.