Electrical tomography is a kind of non-destructive testing technique to image the conductivity distribution within the observation domain. The finite element method is commonly used to solve the inverse problem. The size of the mesh elements can affect the accuracy of the approximation method. The finer size is usually utilized to improve the spatial resolution of the reconstructed image. However, the computational cost will be increased, which makes the inverse problem more underdetermined since the number of unknowns is increased. To address this issue, an adaptive mesh generation method based on the image gradient is proposed to optimize mesh generation to improve the reconstructed accuracy on the premise of not significantly increasing ill-condition. According to the gradient of the initial reconstructed image, the proposed method optimizes the subdivision of the observation field by adaptively improving the mesh density of the inclusion region and reducing the mesh density of other regions. The commonly used mesh generation methods are used to compare with the proposed method. Simulation and experiments show that the reconstructed image error is reduced by 15% on average and the correlation coefficient is increased by 7% on average. Results show that the proposed mesh generation method can improve the reconstruction accuracy of inclusions and the image reconstruction quality, and reduce the calculation error without increasing the number of grids.