As regular inspection for the bolt connection in inaccessible areas is difficult and costly, computer vision technology provides a suitable noncontact approach for real-time bolt looseness detection as an alternative to inspection approaches. However, computer vision still suffers from various impracticalities. In this paper, a new vision-based bolt looseness detection method is designed and implemented with the bolt images acquired by a camera at arbitrary positions around the bolts. The new method includes the perspective transformation of original images acquired, identification of bolt positioning with the convolutional neural network digit recognition, detection of bolt rotation angles using Hough transform line detection, and density-based spatial clustering of applications with noise. To demonstrate the effectiveness of the new method, an experiment with bolted connections is setup. The experimental results demonstrate that the new method can accurately detect the looseness of the bolts in the bolted connection.