Local feature extraction is a key link in realizing computer vision tasks. Although the related research has made great progress, the balance of network efficiency and accuracy has always existed. To address these issues, we propose a lightweight and efficient multi-task feature point detection network. First, an enhanced block (E-Block) is built based on the asymmetric convolution structure to fully mine the feature information, and the parameters of the E-Block are fused by the method of structural re-parameterization to improve the network operation efficiency. Then, we design an enhanced module (E-Model) through E-Block to improve the feature point detection ability of the network. In particular, E-Model utilizes a lightweight shuffle attention (SA) mechanism to reduce redundant feature extraction. In addition, we also design a fast version of the network by exploiting the characteristics of grouped convolution and multi-scale aggregation of pyramid convolution. Experiments on the Hpatches dataset and KITTI dataset show that the proposed network has satisfactory feature extraction ability while reducing the volume. https://github.com/SpiritAshes/LEFPD-Net.git.