Digital Image Enhancement for Improved Diagnostic Abilities
Non-visible light imaging has proven itself useful in many applications. For instance, IR imaging is used in plastic material measurement, condition monitoring or night vision applications. Infrared spectroscopy is used in many on-line process control applications from polyolefin process control to colour measurement of the produced material. UV imaging is used in material inspection, especially in semiconductor industry. X-ray imaging is most prominently used in medical and industrial applications. In order to produce useful results, several image processing methods can be applied when converting non-visible data into visible images. After non-visible image pre-processing it is then easier to use well-known machine vision pattern matching and image measurement algorithms.
In non-visible light images, the pixel values received from a sensor are first converted into a greyscale image. However, direct linear transformation does not produce the best results. There are several reasons for this. First, the sensor itself may be non-linear or produce artefacts, noise, or other non-idealities. Second, the desired information may lie only in a small portion of the received spectrum. Furthermore, there might be different kind of interesting details in different parts of the image. Therefore, the image received from the sensor must be pre-processed before traditional image processing algorithms can be applied. The pre-processing cancels the non-idealities and rearranges the histogram for improved information distribution on the greyscale. After this, the user can apply the traditional image enhancement methods in situ.
Figure 1. Histogram of an image obtained from the sensor and presented in greyscale applying direct linear transformation. The largest amount of pixels is in the darkest end of the spectrum. However, the details are in the other regions.
Figure 2. The image resulting pre-processing of the histogram in Fig.1. The pre-processing included small histogram modification and noise reduction filtering. The histogram modification compensated the sensor non-linearity, which also attenuates the black area noise.
As an example, let us consider X-ray imaging. The pixel values describe the attenuation of the X-ray radiation when it has passed through the subject. The attenuation is exponential with respect to the thickness. Thus, the histogram of a typical X-ray image shows a large peak in the black end of the spectrum (Fig. 1).
The image non-idealities can be cancelled by a suitable pre-processing method. The choice of the method strongly depends on the type of non-idealities, the subject, and the application. For instance, in case of an X-ray image, the exponential attenuation can be compensated by a logarithmic transform of the histogram. The goal of the pre-processing is to produce an image that can be viewed visually. The result is a greyscale image that can be either used as such or processed even further. Figure 2 shows the image, which is the result of pre-processing the histogram in Fig. 1.
After the non-idealities have been removed, the resulting image is a typical greyscale image to which the normal image processing methods can be applied for enhanced diagnostic abilities. These methods include linear and non-linear filtering, spectral transforms (wavelet etc.), and compression algorithms, among others. In Fig. 3, the pre-processed sample image from Fig. 2 has been subject to a common histogram equalisation. The goal has been to make the details more visible. However, this approach also made the remaining noise more visible.
Figure 4. Pseudo-colorization is a very efficient method to highlight a special part of the spectrum. In this case, the shell of the pen is highlighted with blue colour and the brightest parts of the image with red colour.
Figure 5. The histogram of the image in Fig. 6, showing the parts highlighted with blue and red colours.
In order to view details that are in a very small part of the spectrum, pseudo-colorization is an excellent choice. In the example given in Fig. 4, the dense parts are coloured with red colour and the parts that show the shell of the pen are coloured blue, as shown in Fig. 5.
Instead of applying the histogram equalisation to the whole image, it can be applied to a small portion of the image (Region Of Interest, ROI). In Fig. 6, the ballpoint of the pen becomes clearly visible in ROI. The ROI-tools often produce good results when the image is used in visual inspection and the human eye must catch the details fast and accurately.
In application-specific image processing, there are two main issues requiring special attention. First, the technical limitations must be taken into account. This includes such items as sensor non-idealities and X-ray parameters, among others. Typically, in the field of visible light imaging, illumination problems are a similar issue. Second, there must be expertise in the application field in order to find out the points of interest in the images. Thus, one can customize the image manipulation tools not only to solve the technical problems, but also to meet the specific requirements of the application.
The compression of X-ray images usually differs from the compression of the visual light images. The X-ray images typically contain noise that can be filtered out of the compressed image. However, the details to preserve may be very small (small fractures, miniature details), so preservation of sharp edges during the noise filtering is important. In some cases, only a certain part of the image is important. In such a situation, it is possible to compress the interesting part with small compress ratio and the uninteresting parts with high compression ratio. This helps to preserve the details. In some other cases, the details in the border areas are to be preserved only if they differ from the reference (the previous image in a video sequence). This is usually the case when automated detection of exceptional occurrence is desired in industrial monitoring. Furthermore, the dynamic range of X-ray images needs to be considered when compressing the images. If the object to be monitored is made of homogenous substance, only a part of the spectrum contains useful information - the rest may be ignored.
Atostek Ltd has experience in high performance software implementation of computation intensive image processing algorithms on modern PC:s. This is needed especially in X-ray imaging, where images may be large and contain dynamic range of more than 8 bits. In the optimisation process, one should prefer parallel technologies such as SIMD, MMX, and symmetric multiprocessing (SMP). In most cases, the choice of hardware can be left to the user. The platform could be a normal PC workstation, server, blade server, or a PC104-based industrial solution.
Atostek provides software and system development according to the 5-step product development process, the core competence being in image processing, image analysis, and image manipulation. Atostek's services include application analysis producing the requirements of the solution, specification of the proposed solution, feasibility study, architectural design, software implementation, system testing, updates and proposals for further development. Important parts of the software project may include the design of user interface, databases, transfer in IP-network, and integration to other software systems, such as production management systems.
Figure 3. The image in Fig. 2 after histogram equalization. The shell of the pen becomes visible, but the details in the brighter areas are lost. The simple equalization does not produce good results in this case. The stripes occur because of the sensor noise. Due to the noise reduction applied previously (Fig. 2), the noise has been removed from the mid-part of the spectrum.
Figure 6. Local histogram equalisation for region of interest (ROI) reveals details in a hidden in a small part of the picture (the ball-point). The details revealed are in the brightest end of the spectrum.