#TITLE_ALTERNATIVE#

• Main Author: (NIM 23396003), Arifin
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/7287
• Institution: Institut Teknologi Bandung

ABSTRACT: <br /> <br /> <br /> <br /> <br /> This study has been designed an artificial neural network (ANN) structure for geometric correction of raster image data which were collected from the image. This ANN architecture has a supervised learning pattern was that multi-layer perceptron with feed forward network and by means of back-propagation learning rule. This network consists of 3 layers, those are input layer, hidden layer and output layer. <br /> <br /> <br /> <br /> <br /> Geometric correction by means of ANN is examined to ward 40 ground control point of image and map. By using 20 ground control points as learning data, ANN process results error in 1.064 pixel root mean square, whereas for 30 ones it is 0.2045 pixel. As comparative method is used affine transformation and fifth order polynomial methods with 30 control point data. Affme transformation results root mean square in 5.8881 pixel, whereas fifth order polynomial results root mean square error in 3.81 pixel. The results of coordinat transformation by means of both affine and fifth order polynomial methods they have not fulfilled the root mean square error tolerance requirement, maximally 2 pixel. <br /> <br /> <br /> <br /> <br /> Geometric correction process using ANN is better rather than conventional method. For making optimally coordinat transformation model, ANN matchs with the used control point numbers. Whereas both affine transformation and polynomial that transformation model is earlier determined on the basis of order determination is used without setting the available control point numbers. Therefore, by means of control point exceeds the minimum numbers, this process does not make optimally that transformation model but it does only make minimally that model used.