In the cases of appearance changes and viewpoint changes, the accuracy and robustness of traditional visual loop closure detection algorithms become very poor.To overcome this problem, we propose a loop closure detection algorithm, which utilizes the features of multi-level convolutional neural networks. The high-level convolution features contain much semantic information and can cope with viewpoint changes. The medium-level convolutional features contain more geometry and spatial information, which is more robust to lighting changes. Therefore, the accuracy and robustness of loop closure detection is improved by taking full advantage of the characteristics of the middle and high levels convolutional features and modular similarity measures. However, the convolutional feature vectors have a particularly large dimension, so the convolutional feature vectors are firstly dimension-reduced. The experimental results on the Gardens Point dataset show that the image matching detection effect is better by using multi-level convolutional features than by other single layers. In addition, for the dynamic interference factors in the images captured at different moments, a dynamic interference semantic filtering mechanism is further proposed. The filtered images are used to perform the matching. The experiments on the Tokyo 24/7 dataset prove the feasibility and effectiveness of this method.