Experimental Results

The classification of the digital images of cracks is implemented by using the TAM network. Twenty images of cracks are used for learning of TAM network and the remaining 27 images are used for evaluating the classification results. This implementation is repeated many times by changing the learning data every times. In this learning stage, three damage levels judged by an expert for each image are used as the teaching signal. Also, the teaching data include the same number of each damage levels data. In this numerical example, it is evaluated by the recognition rate of non-learning data, and by comparing with the neural network, the effectiveness is examined. The used learning

Table 1. Learning parameters.

Learning factor 0.3

Learning time 100,000

parameters of TAM network and neural network are shown Table 1. In this research, the digital image is divided to 256 blocks, and detected the element of directionally.

The learning parameter is the same as the neural network to compare the performance. Table 2 shows the classification results with the distribution of directionality. This result is average of all trials.

Only about 70% recognition accuracy is obtained by using neural network. Especially, the re­cognition rate of B rank is very low (38.5%). It is caused that same digital image of B rank are close to the A rank or C rank. So, recognition rate of B rank is low. On the other hand, by using TAM network, over 90% recognition accuracy is obtained and the recognition rata of all rank are over 90%. Mainly, the recognition rate of B rank is improvement. From this result, a TAM network can recognize the complex problem that a neural network cannot recognize. It is considered that the proposed system can recognize similar digital image by using TAM network that is modeled ocular system.

Table 2. Recognition accuracy.


Recognition accuracy (%)

A (10 entries)

B (13 entries)

C (24 entries)


Neural network





TAM network






In this paper, an attempt was made to develop a new system that evaluates the damage condition of existing structures by using the visual information given by digital photos. The proposed system is based upon such new technologies as image processing, pattern recognition, and artificial intelli­gence. A new measuring system was developed by using two digital cameras. Moreover, the system for extracting the characteristics of cracks showing up on concrete slabs through digital images was developed and classification based on damage levels was attempted by using these results. First, the linear pattern of cracks is extracted from the digital images of the concrete slabs through image processing techniques. Next, the characteristics such as the projection histograms that are often applied in the field of optical character recognition, and the feature points in the border expression are extracted. Finally, the digital images of cracks are classified into different damage levels based on the extracted characteristics through TAM (Topographic Attentive Mapping) system.