Feature extraction and object recognition aim for classifying objects into a number of categories or classes and the extraction of features and objects from raster data sets.
Extracting information from hard copy maps to develop GIS database represents one of the major bottlenecks in the race against time. The article presents a novel methodology for the extraction and recognition of landmarks and polygons from base maps and images
There is a big demand for digital maps and GIS database in Egypt and all over the world. The production process in Egypt employs manual techniques that are lengthy, time consuming and costly. The article looks at the possibility of developing a semi-automatic module that utilises object recognition and feature extraction algorithms to produce digital maps and GIS database layers from hardcopy maps and VHR satellite images.
The first step in the process is to develop and evaluate a new software module for point & symbol object recognition from raster scanned hardcopy maps and satellite images. Most of the Egyptian maps exist in hardcopy format, including base maps with scales ranging from 1: 25000 to 1: 250000. They are considered to be the primary source of data and are currently transformed to digital format using manual digitising techniques, which are costly and time consuming. The second source of spatial data is satellite imagery. It is useful to develop a software module that can transform the data from raster to vector format and extract various objects in an automatic or semi-automatic procedure. By applying the correlation technique for any image containing many objects, the locations of the desired objects can be automatically located.
Development and evaluation of semi automatic point/symbol feature extraction
To begin the development process, a module is created for feature extraction called the “Symbol Module”, which further consists of four sub-modules. Ten hardcopy maps of scale 1:50000, produced by the Egyptian Survey Authority (ESA) and the Military Survey Department (MSD), are used. These base maps are scanned with resolutions of 200 dpi and 400 dpi, transformed to raster format, and then rectified with projection adjustment. Subsequently, masks for different landmark classes are extracted from the raster maps. The chosen landmark classes are trees, mosques, and schools. One mask is chosen for the trees as they are always found separated on the maps without changes in the background, while numerous masks are selected for the mosques and schools as these features have different backgrounds behind them in the paper maps. There are several techniques that utilise map scanning as a means of recording spatial data. One can scan a map and then use it as a background image for on-screen digitising or use sophisticated software to convert a scanned raster image into useable GIS data. The scanned map cannot be more accurate than the original map. Accuracy refers to how correct the data is, while precision refers to the amount of detail present in the data. Using these rules, it is possible to have a scanned map that is very precise yet inaccurate. One way to control accuracy is to use the best possible maps as a source. A properly controlled environment might not be as important as previously thought.
Symbol Module consists of four sub-modules, which include image browsing sub-module which inputs the high resolution image; Mask browsing sub-module which inputs the landmark that it can be detected; first guess sub-module which makes use of the template matching algorithm to perform primary feature extraction, as well as the enhancement sub-module which determines the threshold to filter and detect the landmarks. Symbol Module uses the template matching technique. A reference template (Mask) is used, and it can be decided where the objects best match with this template. Non-linear spatial filtering and convolution are used through sliding neighborhoods, where the value of an output pixel is equal to the standard deviation of the values of the pixels in the input pixel’s neighborhood. Using one band (red, green, blue) is enough to detect the matched symbols.
Experimental work and results
The first set of experiments is performed for a sub-image for the chosen landmark classes. Three masks (Trees, Mosque, and School) are used. To be successful, four versions of masks for the mosques are used (mosque, mosque1, mosque2, and mosque3).
Taking a big sample from a paper map, different masks for mosques, schools and trees are taken. Using symbol module can extract the mosques, schools and trees. The detected points are bounded by black box. For mosques, the minimum values range from 58 to 79. The maximum values range from 200 to 240. The mean values range from 119.2 to 155.4. Six different templates are used for mosques because mosques have different sizes and different background colors. The number of columns varies from 10 to 14 pixels and number of rows varies from 26 to 33 pixels; for schools, the minimum values range from 72 to 80. The maximum values range from 202 to 241. The mean values range from values range from 121.3 to 156.9. The number of columns varies from 12 to 15 pixels and number of rows varies from 29 to 31 pixels; for trees, the minimum value is 108. The maximum value is 255. The mean value is 204.2. These values are very close to all trees so one mask only for the trees is used.
The experiments are performed for 6 sub-images (A1, A2, A3, A4, A5, and A6) for the chosen landmark classes. Three masks (trees, mosques and schools) are used. For successful experiment, six versions of masks for mosques (mosque, mosque1, mosque2, mosque3, mosque4 and mosque5) are utilised.
After analysing the results, it can be concluded that: The threshold is different for each mask in each landmark; the thresholds of trees range from 45 to 52; the thresholds of mosques range from 27 to 47; and the thresholds of schools range from 20 to 34. That makes the detection and extraction of schools and mosques very difficult. If the threshold is decreased, missed landmarks occur. If the threshold is increased, false alarms occur.
The next step in the process is to develop digital map layers for the recognised features. Firstly, the center of gravity for each matched symbol is defined, which is followed by all the centers (points) being converted into points drawn in a shape file using a point generator module. This module is developed using Arc GIS software and VBA programming language.
In Egypt, maps with scale 1: 50000 consist of more than 14 layers and over 100,000 features (refer to table below). The developed module can be used conveniently to recognise them.
There are numerous advantages of this procedure. It can be used to generate polygons in GIS shape files with the following steps: the procedure creates the centroid of the recognised objects with coding system to classify the objects and save them in a file, using ArcGIS customised application, a vector layer (GIS shape file) automatically using coding system, will be generated.
The above article can be summarized in the below form to draw major conclusions:
- The developed semi-automatic module utilises object recognition and feature extraction techniques for extracting various point, symbol, and area features from raster data sets. It is successful for many reasons, such as reducing time, effort, and cost as well as increasing the accuracy.
- The developed module is color dependent. It can work on color images or gray images. The developed module depends on one band only (red or green or blue). The developed module is scale independent. It can be used with large or small scale maps and with any HR satellite image. Using this module, the thresholds for various feature classes are adjusted in the beginning of the work and then, the same thresholds are used for all other images and maps from the same data set.
- The developed module can recognise the images with random error values ranging from 0 to ±140 and systematic error values ranging from -80 to +80
- Since only one band is used in the recognition system, the shape of the mask should be taken from a typical image of the data set. Other shapes in the same data sets could be identified and extracted if they are similar to or close to the mask image. To recognise the same mask, the same background in the image must exist.
- The developed module has many advantages over other similar existing systems. Some systems are developed for either maps or satellite images while the developed module can process both of them. Some systems use only monochrome or color images, while the developed module uses both of them. Some systems do not have known minimum area to be recognised, while the developed module can recognise a minimum area of 13 x 13 pixels. The developed module uses a multiple object library which is the novel technique used in object recognition and hence, the output vector features are smooth and clean to be used directly in the GIS database without a need for extra editing.
- Four main properties should exist to have successful area recognition: matching colors, identical shape, same background, and as rectangular as possible.