Seismic inversion, a critical method for inferring subsurface petrophysical parameters (acoustic impedance) from seismic data, plays a vital role in hydrocarbon exploration and geological research. However, conventional seismic inversion methods heavily rely on low-frequency background models, exhibit sensitivity to noise interference, and suffer from low computational efficiency, making them inadequate for high-precision inversion under complex geological conditions.In recent years, Convolutional Neural Networks (CNN) have been widely adopted in seismic impedance inversion due to their strong nonlinear modeling capabilities and end-to-end learning approaches, enabling effective mapping of seismic data to impedance values. Nevertheless, deep CNN architecture still face challenges such as gradient vanishing, gradient explosion, and overfitting, which compromise the stability and accuracy of inversion results. To address these limitations, this study proposes a ResUNet-based inversion approach by integrating residual blocks into a UNet framework, thereby mitigating gradient vanishing and information loss issues while enhancing inversion accuracy and stability. Synthetic seismic data generated through forward modeling were used to train the ResUNet model, and real seismic data were subsequently input into the trained network to derive impedance inversion results. However, the intelligent inversion outputs may exhibit discrete artifacts and "spiking" phenomena near complex geological features (fault), impairing the smoothness and continuity of the results. To optimize the inversion outputs, a post-processing smoothing technique was further introduced to suppress anomalies and improve stability. Theoretical model testing and real-data validation demonstrate that the ResUNet combined with smoothing effectively enhances the precision and stability of seismic acoustic impedance inversion, offering a feasible novel method for intelligent seismic inversion applications.