Self-piercing rivet (SPR), a method used for joining sheet materials, is of increasing interest in the automobile industry. Cracks on the SPR joint surface might affect the joint strength significantly. This paper proposed a novel crack detection and evaluation method for SPR button images based on machine learning, which will address the issue of time-consuming and subjective caused by manual visual crack inspection. Firstly, sub-images are cropped from the button images and preprocessed into three categories (i.e., cracks, edges, and smooth regions) as training samples. At the sub-images level, 5 neural networks are trained with the input of different extracted features, respectively. Then, to overcome the representation limitation of one single extracted feature, a weighted combination of 5 neural networks is developed. Thirdly, a search algorithm is developed to extend the application of the learned model from sub-images into the original button images. Lastly, an evaluation system based on the characteristics of SPR button images is proposed to compare and analyze these different application results. The preliminary results on non-cracked and cracked button images show that the proposed crack detection method of weighted combination is an effective approach, which gets better crack detection results compared to other crack detection models with a single extracted feature of input.