The disclosure discloses an online detection method of circular weft knitting stripe defects based on a gray gradient method, and belongs to the technical field of textile product detection. The method provides a defect detection and positioning method based on the gray gradient method. Before the detection on a new product, only a model needs to be trained to obtain stripe defect feature images with different stitch types and different stitch densities, and the detection is directly performed in a subsequent process. Defects can be fast and accurately recognized, and a cam position causing the stripe defects can be calculated according to a quantity of shot images between the feature images and a marked image, courses in which marks in the marked image are located, courses in which defects in a defect image are located, and machine operation parameters. Moreover, a defect detection device required by the disclosure can be modified on an original circular knitting machine, so the detection cost is reduced.