Object localization in 2D images based on Kohonen's self-organization feature maps
Chuan Yuan, Hendrik J. Niemann · 2003
This paper presents a hybrid approach for neural object localization and recognition in 2D grey level images. The system combines an auto-associative network, two self-organization feature maps (SOM), and a three layer feedforward network trained with dynamic learning vector quantization (DLVQ). By using a hidden layer smaller than the input/output layers, the auto-associative network can be expected to find efficient ways of encoding the information contained in the input data set. Thus a dimension reduction of the input image can be achieved. The object localization scheme is then directly based on features which are detected automatically using the Kohonen's SOMs. After preprocessing images are split into small blocks and input to two Kohonen maps. Through training, the first map can detect the object area of the input image, while the second map can detect the object specific features. By integrating the features extracted from the output of the two maps and the DLVQ methods, we can locate different objects and estimate object pose (translation, rotation within the image plane and scale parameter).