This paper mentioned the fingerprinting/model matching case. But I could not find an image based algorithm. Any suggestion about image based localization
I'm not sure what you mean by "image based," but it seems like the term "model matching" is used to describe the notion of using an a-priori map of an environment to localize to.
You can build an a-priori map in many ways such as using a lidar, or using detected features from an image. Once you know this global map, your robot can use its sensor data to position itself on the global map by using localization algorithms such as a particle filter or EKF.