Development and in-orbit demostration of a realtime image recognizer and its application to the world’s first 3-axis attitude determination is presented. To deal with the limited computational resources and power supply, we have developed an image recogniz- ing algorithm utilizing a multi-stage neural network. This employes the color information as a feature quan- tity rather than the molophology to reduce the com- putational cost. After the training with the in-orbit data, the discrimination accuracy finally achieved to 78%, a level close to the human eyes. Furthermore, by comparing the land shape extracted by the image rec- ognition with the map information, the 3-axis attitude determination was demonstrated. Above the coastal area with good weather condition, the algorithm can obtain 3-axis attitude with an accuracy of ~3°.