Self-supervised learning (SSL) is a process of learning a general-purpose feature extractor using artificially generated label data. Depending on the application, invariance and equivariance are desired in image feature representation. Invariance is a property in which feature representation does not change before and after applying transformations such as image scaling and rotation to an input image. In contrast, equivariance refers to the property in which feature representations change in response to the transformation of the input image.In SSL, whether invariance or equivariance for a given transformation considerably affect the downstream target task. However, a method for measurement of equivariance is not yet be proposed. In this study, inspired by observation of feature expressions and canonical correlation analysis, a novel metric for equivariance measurement was proposed. We used the MNIST, fashion-MNIST, and CIFAR-10 datasets to verify the behavior of the proposed metric of same-degeneration for encoders learned in a rotation prediction task, a type of SSL.