Indoor positioning methods based on CSI (channel state information) of WiFi (wireless fidelity) have extensive application demands. However, in noisy environments, the effective features of WiFi signals can be obscured, leading to a decrease in positioning accuracy. To address this issue, an indoor positioning method based on DRSN (deep residual shrinkage network) was proposed. This network introduces soft thresholding in the nonlinear transformation layer to eliminate unimportant features. To enhance the network's learning ability, two threshold learning channels were set up, which respectively learn the amplitude difference threshold and phase threshold for each antenna and subcarrier of the fingerprint, and it was named DRSN-AS (deep residual shrinkage network with antennas-wise and subcarriers-wise thresholds). In the offline stage, for different experimental scenarios, the amplitude difference and phase information were weighted and combined, and then sent to the model for training. Experimental results show that the positioning accuracy can reach 1.54 m in the self-study room scenario and 0.99 m in the corridor scenario.