Abstract State of charge (SOC) estimation constitutes a critical task of battery management systems. Conventional SOC estimation methods designed for dynamic profiles have difficulties in estimating SOC for LiFePO4 batteries due to their flat open circuit voltage characteristics in the middle range of SOC. In this study, a deep neural network (DNN) based method is proposed to estimate SOC with only 10-min charging voltage and current data as the input. This method enables fast and accurate SOC estimation with an error of less than 2.03% over the entire battery SOC range. Thus, it can be used to calibrate the SOC estimation for the Ampere-hour counting method. We also demonstrate that by incorporating the DNN into a Kalman filter, the robustness of SOC estimation against random noises and error spikes can be improved. In the case of significant disturbances, the method still maintains a root mean square error of 0.385%. Moreover, the trained DNN can quickly adapt to various scenarios, including different ageing states and battery types charged at different rates, thanks to the transfer learning nature. Compared with developing a new DNN, transfer learning can provide more accurate estimation results at less training costs. By only fine-tuning one layer of the pre-trained DNN, the root mean square error can be less than 3.146% and 2.315% for aged batteries and different battery types, respectively. When more layers are fine-tuned, superior performance can be achieved.