SLAM estimates the pose of robots and reconstruct ambient environments simultaneously, which plays an important role in autonomous exploration. The low-light and low-texture challenges dramatically decline its performance, where neural networks are usually adopted for better performance. However, many existing works focus on performance with heavy computation, resulting in the infeasibility for real-time deployment on resource-limited edge platforms. To solve the problem, we propose tailored neural networks and switching strategy for high computational efficiency. First, we propose the illumination enhancement network based on previous work and adopt the GCNv2 for robust feature extraction. Second, we design the dynamic switching strategy based on quantitative evaluation of challenges to automatically turn on and off networks according to different conditions. Finally, we integrate networks and the switching module to ORB-SLAM2 in a modular way and develop a visual SLAM. Extensive experiments on EuRoC and field-collected sequences demonstrate the feasibility of our method with high performance and efficiency.