| Summary: | 碩士 === 國防大學理工學院 === 機械工程碩士班 === 101 === Nowadays, the military ammunition identification work is still artificial manner. Therefore, this study attempted to establish recognition system through image processing technology combined with neural network to improve the recognition results of ammunition types. The results of recognition affects mainly by artillery shell features, shape and color characteristics identify the most effect.
First, to extract the shape feature of ammunition images, modal analysis and some image intervals developed method are applied and the connected component is used to implement shape reconstruction. And to avoid identification being affected by the position, size and rotation angle, it is necessary for shape reconstruction graphics to be normalized. Finally, the geometric invariant moments of the normalized graphics are as the shape features.
For color feature extraction, after constructiing back-propagation neural network color classifier by color samples, each pixel color of the artillery shell graphics is classified and the area ratio of each color is calculated. Shape features and area ratio of each color are used to construct two kinds of shells neural network recognition system, its network parameters are adjusted through the constructive algorithm to get the best recognition system.
Considering the recognition results would be affected by environmental disturbances, i.e. camera angles, shooting distance and kinds of light, in this case, a large number of samples are generated by three degrees of these factors. Orthogonal array is used to diminish training samples and the results demonstrate that training time is reduced effectively at each iteration and keep excellent recognition capability. On the other hand, although the changes of light and color cause the incomplete developing shape, its geometric invariant moments can be still effectively identified. By contrast, the color characteristics except are affected by the incomplete shape, color distortion is more severe disruption. Hence, shape recognition system has better identification effect and the appropriate input values bring up a desired recognition system.
|
|---|
