Frequency estimation is a key technology in the field of measurement instrumentation. Traditional periodogram and multiple signal classification (MUSIC) algorithms suffer from severe performance degradation under low signal-to-noise ratio environments, failing to meet the precision requirements of instrumentation measurements. This paper proposes a frequency estimation method based on multi-scale residual networks (MultiFR). First, parallel 3×1 and 5×1 convolutional kernels extract features at different scales, enhancing the ability to capture both local details and global patterns. Then, residual connections mechanisms are introduced to overcome the gradient vanishing problem in deep networks, ensuring effective feature propagation. The L2 norm squared loss function is adopted to achieve precise measurement of the deviation between the network output spectrum and the ideal spectrum. Finally, a dynamic SNR training strategy is utilized to improve the model’s adaptability under different noise levels, and a high-resolution frequency representation output module is implemented to achieve accurate frequency peak localization. Experimental results demonstrate that the proposed method achieves improved frequency resolution capability. Under extremely low SNR conditions of 0, MultiFR achieves a miss detection rate of 22.2% and a Chamfer distance of 0.145, representing relative improvements of 4.1% and 5.8% respectively compared to the DeepFreq method, and outperforming mainstream methods such as PSnet, MUSIC, and periodogram. Validation with Coriolis flowmeter signals confirms the feasibility, effectiveness, and robustness of the proposed method under low SNR conditions.