Current feature detection in particle flow fields was commonly measured by the particle image velocimetry （PIV） followed by manual identification of multiple consecutive frames, which is subject to some subjective errors. Therefore, a DC-UNet++ network for identifying the quasi-static regions in single-frame flow field images was proposed. Firstly, the particle flow field images formed by the impact experiments between the small ball and particle bed were captured by a high-speed charge coupled device （CCD）camera. Next, the raw images were used for PIV analysis and a dataset was produced. Then, the CNN model, UNet++ model and the proposed DC-UNet++ model were trained on multiple datasets to validate and analyze their feasibility and accuracy in detecting the quasi-static regions on a single frame image. Finally, the ability of the model to generalize over low velocity impact flow fields in non-transparent and transparent granular materials was discussed. The experimental results showed that the DC-UNet++ network achieved an accuracy of 87.76% and 72.91% on non-transparent and transparent granular materials, respectively. The DC-UNet++ network achieved the task of detecting the target features on a single image frame and still had a relatively accurate detection of features under complex flow fields in transparent granular materials.