Abstract: When the image is compressed adaptively with compressed sensing theory, the determination of sampling rate and sparsity threshold were highly subjective. In order to solve the problem, an accurately adaptive sampling algorithm with sparsity fitting was proposed in this paper. This algorithm determines the minimum sampling rate under certain sparseness to meet the PSNR requirements by iteration, and an optimal objective function of sparsity-sampling rate choices was obtained with the method of least squares fitting sparsity and sampling rate data. The adaptive sampling algorithm was simulated based on TVAL3. Experimental results show that the PSNR values of reconstructed images are higher than that with the same fixed sampling rate algorithm, and the PSNR difference of clear texture distinction images can reach more than 3.5 dB. Compared to the roughly adaptive algorithm, when the average sampling rate is lower than that, the reconstructed image obtains a higher PSNR value.