The traditional visual SLAM system often encounters challenges such as low pose estimation accuracy and tracking failures, especially in scenarios with variations in viewpoint and illumination. While deep neural networks, excels in image feature extraction, its high computational complexity and slow runtime speed remain significant drawbacks. Therefore, this work proposes a lightweight feature extraction network-based visual SLAM algorithm. Firstly, to mitigate the reduced real-time performance caused by the heavy computational load of the SuperPoint network's feature encoder layer, we employ a lightweight Encoder network based on GhostNet to alleviate the computational burden. Secondly, in terms of network training, we introduce a novel knowledge distillation method to transfer knowledge from the original SuperPoint network to the new lightweight network, thereby enhancing the network's robustness. Finally, we integrate the lightweight feature extraction network with the backend of the SLAM system, thereby forming a complete monocular visual SLAM system. In comparison to SuperPoint based SLAM, our proposed framework achieves a fourfold increase in runtime per frame, while also surpassing SuperPoint-based SLAM and reducing trajectory errors by approximately 30% compared to ORB-SLAM2.