In modern complex electromagnetic environments, the received radar signals are often affected by multipath effects. The distortion and noise of the multipath signals make it difficult to recognize the intra-pulse modulation of radar signals. To address this issue, a gram angular field-fusion Inception-Res2Net with squeeze-and-excitation (GFSE-InRes2Net) is proposed. Firstly, in addition to the conventional input of time-frequency domain, gram angular field (GAF) is utilized to convert the multipath radar signals into the polar coordinate domain. This allows for the extraction of additional information regarding the multipath signals, such as angle and phase. Then, within the Inception serving as a multi-scale feature extraction module, the Res2 basic block is employed in lieu of traditional convolution operations, allowing for the representation of features at a granular level. Finally, the squeeze-and-excitation (SE) block is used to weight feature channels, which mitigate the impact of multipath features. The comparison and ablation experiments demonstrate the effectiveness of the proposed method.