The nondestructive inspection of the welded region is very important for quality control. The purpose of this study is to develop a methodology for nondestructive evaluation of the function of the spot-welded joint. In the previous paper, the system that consist of an ultrasonic inspection device and neural network was proposed to evaluate the tension shearing load. The welded region's geometry in resistance spot welding of a two-plate joint was inspected by parallel scanning over the region by ultrasonic inspection. The neural network learned the pattern sets dealing with the interaction between the scanning graphs and the tension shearing load of spot-welded joints. The validity of the previous system was confirmed by comparing the output of the network with the experimental results. However, it had been serious problems for improving the accuracy of the tension shearing load that the learning efficiency of the neural network was not good and the internal structure of the network was involved (the number of the input layer's units was 2401 units). In this paper, fractal dimension is introduced to characterize the scanning graphs, and the application of the dimension is proposed as the input data of the neural network. The following became clear after the investigation : (1) The scanning graphs showed fractal characteristics. (2) The number of the input layer's units was reduced from 2401 to 98. (3) The accuracy of the evaluation and the learning efficiency of this network were much better than the previous network.