In order to solve the problem of accuracy of inversion of corn canopy LAI, it is necessary to study the effect of specular reflection on the image reflectance of unmanned aerial vehicle (UAV), which is independent of canopy structure. The wavelet transform was used to set the threshold of different bands of UAV image, and the specular reflection was weakened without affecting the diffuse reflection. The vegetation indices: NDVI, GNDVI, SAVI and EVI were constructed by using multispectral UAV images of the Hebei Agricultural University Xinji Test Station acquired on July 15th and 26th, 2018. The single-variable inversion model of maize canopy LAI was constructed, and the accuracy of LAI inversion was evaluated by R2 and RMSE. The results showed that when the maize plants were sparse on July 15th, the R2 of vegetation indices and measured LAI after removing specular reflection were raised from 0.7190, 0.5598, 0.6241 and 0.5985 to 0.7633, 0.6940, 0.6497 and 0.6194, and the RMSE was also decreased from 0.2244, 0.2526, 0.2214 and 0.2245 to 0.1880, 0.1958, 0.1918 and 0.1987, which showed that removing specular reflection can improve the accuracy of LAI inversion. On July 26th, when the maize plants were relatively dense, the R2 of the four indices were also increased after the removal of specular reflection, which proved that the removal of specular reflection could improve the correlation between vegetation indices and LAI. However, in this case, NDVI and GNDVI tended to be saturated, and reducing the reflectivity by threshold method would aggravate the saturation phenomenon, so the two indices could not fully reflect the change of LAI. Meanwhile, SAVI and EVI were amplified by adding a canopy background adjustment factor, and their R2 of fitting model with ln(LAI) were both over 0.6 after removing specular reflection. Thus SAVI and EVI were more suitable for LAI inversion when vegetation cover was dense.