|
|
|
Digital Image Processing
|
The use of image processing systems for the analysis of digitised Aerial Photography
Method of data input
Analogue photographs can be digitised into processing system using a variety of techniques. Two every common techniques are:
- Frame Grabbing:
The simplest method producing raster data is by frame grabbing the imagery . This is a low cost option because much of the equipment is readily available. To frame grab an image requires a video camera connected to a personal computer. In this particular application a standard video camera was connected to a special board within the personal computer. This board frame grabs the analogue data and converts it into 8 bit in a 512 by 512 pixel image. This image can then the easily imported into an image analysis system.
- Electronic Scanning
There are a wide range of scanners on the market that are able to digitise aerial photographs, included with many of these scanners are software programs that allow zooming, basic filtering and printing of the imported images.
For this particular application, additional software items were not required, since the image could easily be manipulated within the image processing system.
The technical specifications of the majority of scanners on the market are similar, varying mainly in accepted sheet size and scanning resolution. For an excellent summary of the types and specifications of image scanners see Bosma et al (1989) and Bosma and Drummond (1989).
A flatbed scanner was used which scans at a rate 25 milliseconds per line and provides a maximum resolution of 300 dots per inch (dpi) . Images are scanned in 8 bit or 24 bit resolution using a charge coupled device (CCD) With 8 bit resolution 256 levels of grey can be scanned and displayed in monochrome. With the 24 bit resolution the full 16.8 million colours are available, however, the 24 bit mode requires high resolution video adaptors and monitors on the computer system. (Harrison, 1990). Both black and white (one pass) and colour separations are possible (multiple passes using a set of red, green and blue fluorescent lights).
Image processing
- Data quality:
Once the scanned image is in a digital form, image enhancement techniques can be applied to the data.
Linear contrast stretches where initially performed. The images exhibited a large dynamic range and so were particularly amenable to contrast enhancement.
In order to assess the quality of the data, principal component transformations were carried out. These were computed in the first instance on the Flinders Ranges data (multi spectral). From these the noise to signal ration (NSR) were calculated for both the scanned and the frame grabbed imagery.
The Noise to Signal ratio (NSR) is a statistic calculated by the principal component analysis algorithm which summarises the variance of a component relative to the whole data set (Harrision and Jupp, 1990). Typically, the ratio of noise to signal ratio for remotely sensed data should be low at around 5-10 %.
- Frame Grabbed Imagery:
A NSR of 45% was computed for the grabbed imagery. This was expected as the data exhibited a large amount of periodic noise when viewed on the screen.
In an attempt to 'clean ' the frame grabbed imagery a median filter was applied to the raw data. This was unsuccessful due to the dominating nature of the periodic noise. Examination of principal component (PC) 2 revealed that this band contained the majority of the noise. Therefore, the median filter was applied to this but was also unsuccessful. A median filter applied to PC 1 produced an acceptable result. A second pass using a median filter removed the reminder of the noise resulting in an improvement in PC1. The second and third PC's were cleaned after applying the median filter a second time, however the still substantial amount of noise made both components unusable.
The same procedure of noise determination we repeated on the colour Philip Island data, again showing in high NSR of 28% and using principal component analysis revealed large noise factor in components 2 and 3 . The 1st component , again as expected, produced the 'cleanest' image.
All data sets showed similar NSR's and it was therefore decided that the data was of unacceptable quality and no further processing was undertaken. As the 2nd principal component shows (see figure 2) the noise throughout the image has a regular pattern. As a result it may regular pattern. As a result it may be that a Fast Fourier Transform (FFT) algorithm could be applied to the data to try and reduce this noise.
Figure 2: 2nd principal component Flinders Ranges.
|
|
|
|
|
|
|
