Aircraft image recognition using back-propagation

Aircraft Recognition Based on Isodensity and Ordinary Moments

“…It is apparent from Fig. 6 that it is still the aircraft identification system based on the 2-D combination moments defined by equations (5) and (6) is superior than that one using the first two components of the invariant moments set.…”

“…In this paper, the following combinations of the m pq calculated using individual isodensity regions have been used to form a feature vector of each aircraft for recognition purposes. Since equation (5) represents the area of each isodensity region, and equation (6) contains the same combination as equation (3), equations (5) and (6) are invariant to the translation and rotation (if the rotation is in x-y plane). Scale invariance can be achieved easily, if the clipped aircraft images are normalized using the same manner like Khotanzad method [7].…”

“…It was found that the identification failures were usually associated with boundary variations in the aircraft image of a given model due to changes in the orientation between settings. The second method was carried out after replacing the combined moments defined by equations (5) and (6) with a 2-D invariant moments defined by equations (3) and (4) and the recognition performance is illustrated in the second row of the Table 1. It is apparent from Table 1 that the best recognition rate achievable was 85.7% with L = 8 which is significantly worse than those presented in Table 1 (row 1) as shown in Fig.…”

“…Other recognition systems use the grey-scale digitized image of an aircraft directly without pre-processing operations such as segmentation and these include some well known systems such as WISARD, back propagation networks and Kohonen's matrix system, all of which are reviewed in reference [4]. Somaie et al used the back propagation network to recognize aircrafts [5]. In this paper, the presented training algorithms satisfied high recognition performance when the number of hidden neurons was less than or equal ten times the number of classes.…”

“…In this paper, the presented training algorithms satisfied high recognition performance when the number of hidden neurons was less than or equal ten times the number of classes. It was found that the network recognizes the test images correctly even the noisy and the incomplete images [5]. An information theory approach has been exploited by Somaie et al [6], who used eigencrafts (in a low dimensional space) to model aircrafts and to distinguish them in a limited database of aircrafts.…”