Aiming at problems such as long cycle and insufficient real time information in traditional remote sensing ground target image recognition system, an on-board target fast matching recognition platform is designed for fast on-orbit satellite recognition, and an improved feature matching recognition algorithm based on fast retinal key points (FREAK) is proposed to solve the problems of complex backgrounds and large amount of data in remote sensing image, First, we introduce the principle of on-board target recognition system and propose the simplified FREAK feature extraction model, and then we reduce the model of original algorithm from seven floors to four to quickly extract target features in remote sensing image. And then the high-dimensional feature data is quantified into two-dimensional data using binary quantization space, thus improving the accuracy of the algorithm; finally, the remote targets are recognized quickly by matching. The experimental results show that the matching accuracy can be increased by 2.3%, and matching time can be reduced by 27.8%. It can meet the requirements of quick identification of remote sensing satellite on-orbit targets.