We have developed a system to acquire gear side-images at specified time intervals during durability tests of plastic gears with a high-speed camera, and proposed a method for quantifying the crack length at tooth roots. However, we did not consider the influence of the shooting angle of the high-speed camera, and the evaluation method has a possibility that the quantitative value of cracks depends on the current shooting angle. Therefore, in this study, we performed the homography transformation on the side-images of gears taken from an arbitrary position to correct the images from the normal direction. Then, we evaluated the influence of the shooting angle by comparison with the anteroposterior relationship of the transformation. Furthermore, we tried to classify automatically the level of crack length at tooth roots from the captured images by a Convolutional Neural Network (CNN) system. As a result, the influence of the shooting angle was limited, and the CNN system should be effective for the classification of the crack level.