The intelligent reconstruction methods for supersonic flow fields face challenges such as the loss of complex wave system structure features and the inability to effectively capture the temporal evolution characteristics of unsteady flow fields, leading to inaccurate identification of the shock train leading edge. A neural network model based on combined detail feature enhancement to address these issues was proposed. High-precision predictions of the density gradient field was achieved based on sparse pressure data. The main wave system structure features of the flow field was established by connecting multiple layers of convolutional networks in series. A residual network with skip connections was used to integrate features from receptive fields of different scales, enhancing the model's ability to express detail features in reconstructed flow fields. Validation was conducted using a dataset constructed from numerical simulations of ramjet engines. Compared to multilayer convolutional neural networks, this method improves the average Peak Signal-to-Noise Ratio across the entire test set by 9.5%. Moreover, the reconstructed flow field's STLE position closely matches the numerical computation results, further demonstrating the effectiveness of the proposed method.