Minimum average correlation energy algorithm (MACE) used in matched filter is improved and applied to real-time joint transform correlator. A set of spectram of training images are synthesized with MACE and synthesized reference images are got after inverse Fourier transform. Meanwhile,the edge extraction of training images is made before reference images synthesis,Laplace sharpening to the joint power spectrum of object image and reference image are carried out in order to enhance contrast of correlation peak. Rotation variant object images and optical experiments are showed. The experiment proves that improved MACE method can decrease training image number greatly,reduce calculation and well solve the problem of rotation variant object recognition in joint transform correlator. Target recognition scope is expanded and recognition rate is increased.