To deal with the mismatching between elevation and azimuth angles in orientation angle estimation of coherence signals by 2-D Estimating Signal Parameter via Rotational Invariance Techniques (ESPRIT)-like algorithm designed for the two-dimensional cross-shaped MEMS (Micro-Electro-Mechanical System) ultrasonic phased array, the improved 2-D ESPRIT-like algorithm based on the joint diagonalization of receive signal matrices was proposed. Firstly, the correlation matrices along x and y axis were derived by the received signal matrices to reconstruct the Toeplitz matrices to decoherence correspondingly according to ESPRIT-like algorithm. Secondly, the Toeplitz matrices were decomposed equivalently to obtain the equivalent received signal matrices after decoherencing. Finally, the joint diagonalization was used to diagonalize the equivalent received signal matrices to realize the matching between elevation and azimuth angles and estimate the orientation angles correctly. The simulation results show that the improved algorithm can estimate the orientation angles correctly compared with the algorithm before being improved. In comparison with the commonly used 2-D MUltiple SIgnal Classification (MUSIC) algorithm based on spatial smoothing, the response time of the proposed algorithm is decreased by 79%, the resolution of elevation and azimuth angles is increased by about 20% and 40% respectively, and the angle error is about 10% that of MUSIC algorithm when the SNR is 30 dB.