Addressing the issues of insufficient accuracy and low computational efficiency in quantitative precipitation estimation for dual-polarization radar, this paper proposes a lightweight multi-scale fusion algorithm, DPCR-Net, based on an improved DeepLabV3+. This algorithm employs a lightweight backbone network combined with inverted residual and dilated convolution to reduce parameters while maintaining feature extraction capabilities; enhances feature reuse and sensitivity to small targets through a densely cascaded adaptive feature pyramid; and utilizes a hybrid attention decoder to fuse multi-scale features, suppress noise and preserve details. Experimental results on the NJU-CPOL dataset show that the method achieves a mean absolute error (MAE), root mean squared error (RMSE) and correlation coefficient (CC) of 0.519 7, 4.174 4 and 0.661 0, respectively, with a hit rate of 0.897 4, a parameter count of only 4.11 M and a computational cost of 8.29 G. Compared to mainstream models, the algorithm proposed in this paper improves both estimation accuracy and computational efficiency, enabling high-precision real-time precipitation estimation and deployment on edge devices.