The intrapulse modulation radar signal detection is essential for contemporary electronic intelligence reconnaissance and other domains. However, there are difficulties in modulated signal recognition due to the complicated electromagnetic environment and poor signal-to-noise ratio. Even though the recognition accuracy has been increased by the current methodologies, the computing pressure are high, and generalization capacity are limited either. In this paper, the Dual-Path Dilated Convolution Attention (DDCA) is proposed. This method is proposed as a lightweight improvement technique to increase the model performance, and a lightweight recognition method by combining denoising convolutional neural networks and improved MobileViT. The simulation results demonstrate that the proposed method effectively reduces the parameter size of network. Meanwhile, a sustainable high recognition accuracy is obtained, the network inference speed is accelerated, and the hardware requirements are alleviated. Based on simulation, the recognition accuracy for ten types of radar signals attains as high as 91.4% when the signal-to-noise ratio is - 16 dB.