A method and apparatus for encoding and decoding an image by performing motion prediction and compensation on pictures in a group of pictures by selectively using a high-quality key picture that is previously encoded and restored, and a second picture that is previously encoded and restored. The method of encoding an image improves the prediction efficiency of an image by storing a key picture, which is encoded and restored to a high quality by reducing a loss caused by a quantization error using a small quantization coefficient, in a first storage unit and storing a previously encoded and restored second picture in a second storage unit, and then performing motion prediction and compensation by selectively using the key picture stored in the first storage unit and the second picture stored in the second storage unit while encoding a next picture in a group of pictures.