期刊名称：Archiwum Fotogrametrii, Kartografii i Teledetekcji
印刷版ISSN：2083-2214
出版年度：2007
卷号：17
语种：English
出版社：Main Board of Association of Polish Surveyors
摘要：This paper describes the application of neural networks for selection of sub-images and the result of the search for the selected areas on the remaining photographs with the utilisation of Kohonen’s SOM network responses distribution. Image fragment representation based on the gradient magnitude values distribution and its direction was proposed. The research was conducted on nine hundred sub-images,taken from aerial photographs of the Cracow’s environs with different terrain cover,divided into three categories: advantageous,intermediate and disadvantageous areas in respect of searching for the features for mutual matching. The edges were detected with Canny algorithm. Based on the gradient values and the directions of the edges,the histogram was created and used to determine the representation of the sub-image in the direction’s profile form. The prepared representation served for teaching the neural network using supervised (backpropagation) and unsupervised (Kohonen) method and later for the classification. For the backpropagation network,the classification effectiveness was measured using the global recognition coefficient and the cooccurrence matrix. For the Kohonen network,the completeness and correctness coefficients were determined. Afterwards,the two networks were put together: the responses generated on the Kohonen map constituted the input signal for the backpropagation layer. In the next step,the adjacent images were sought for the chosen areas. Response distribution on the Kohonen network feature map, amplified with the function preferring the high values,was compared with the distribution for the same size sub-images of the adjacent photographs. To measure the similarity of the subimages,the correlation coefficient to compare network’s responses was used.
其他摘要：Artykuł przedstawia wyniki wykorzystania sieci neuronowych do selekcji podobrazów oraz wyniki wyszukiwania wybranych obszarów na pozostałych zdjęciach z wykorzystaniem rozkładu odpowiedzi dla sieci SOM Kohonena. Zaproponowano reprezentację fragmentu obraz
