Feature detection is a basic problem in computer vision. The ability of feature detectors to detect and extract features directly influences the effect of subsequent image processing. Currently available hand-crafted detectors are poor in extracting repeatable features, and some learning-based detectors have complex network structures and slow detection speeds. In this study, a lightweight network feature detector VGG(visual geometry group network)-Det based on deep learning is proposed to deal with these problems. Through the inter-layer reduction and structural adjustment of the traditional VGG16 network model, a compact network model is designed. Randomly transformed standard image patches are used as the input to train the network to learn image information with high efficiency and extract local features rapidly. In this experiment, the performance of VGG-Det on the three classic datasets of Webcam, EF(edge foci), and VGG-Affine is evaluated, and average repeatabilities of 67.9%, 66.9% and 47.1% are achieved, respectively. VGG-Det has a repeatability of 11.4% higher than that of the newer TILDE-P24 detector. In terms of time performance, VGG-Det has a significant advantage over TILDE-P24 with an equivalent repeatability. The detection speed is increased by nearly 21.5%. Experimental results and principal analysis show that VGG-Det is efficient and feasible. It synchronously improves the speed and repeatability of image feature detection.