The invention discloses a method for detecting panoramic image registration effects, belongs to the technical field of image stitching and image quality evaluation, and relates to the field of computer vision. The method includes the following steps: S1. generating a training data set and a test data set; S2. generating a convolution neural network model; and S3. training the test data set on the basis of the trained convolution neural network. According to the invention, the method, based on an output label of each block output by the convolution neural network, computes the average value of the output labels of all the blocks which constitute a to-be-evaluated stitched image, then computing the average value of the output labels of all stitched images which are synthesized with the same stitching algorithm, the average value of the output labels of all the stitched images serving as a criterion for evaluating the effects of the stitching algorithm. The method, by means of the convolution neural network, can replace tedious and bulk amount of manual statistics and scoring, can accurately determine the clarity of a fusion region in stitching images, overcomes the limitation brought by single-factor evaluation index, can help the realization of automatic and self-adaptive image stitching system, and has significant application value.