We proposed a femtosecond laser reconstruction method based on a multi-output residual neural network. Using this method, we performed quality analysis and pulse inversion on the trace of frequency-resolved optical gating method. Furthermore, we optimized the inversion results using a local weighted regression method. Results show that the trace quality recognition model in the preprocessing stage of proposed algorithm achieves an accuracy of 98.14%. Compared with retrieved amplitude N-grid algorithmic (RANA), the proposed algorithm's reconstruction result has an average relative error of ~4.6%. The average calculation time of the proposed algorithm is ~0.037 s, indicating that the calculation speed is more than an order of magnitude faster than that of the RANA. Additionally, the proposed algorithm has strong noise immunity, demontrating the feasibility of the residual neural network in femtosecond pulse inversion. This method is important for the rapid reconstruction of femtosecond pulse lasers and improving stability at low signal-to-noise ratios.