Loop closure detection algorithm is essential for the visual simultaneous localization and mapping (VSLAM) systems to reduce accumulative error and build a globally consistent map. When detecting loops under the change of viewpoint and scene appearance, the precision and robustness of traditional loop closure detection algorithms decline and some algorithms based on deep learning are difficult to extract features and perform loop closure detection in real time. To overcome these problems, we propose a novel loop closure detection algorithm based on convolutional autoencoder fused with Gist feature, forcing the encoder to reconstruct the Gist feature to enhance the expressive ability of the model when the scene appearance changes. In the same time, we warp images with randomized projective transformations to make the training pairs to improve the precision and robustness of the model when the viewpoint changes. Our model is relatively lightweight which is capable of extracting keyframe features and detecting loops in real time. The results of experiments on Gardens Point and Nordland datasets show that our model can achieve better precision and robustness compared with traditional methods, like bag of visual word (BoVW), Gist, and some other methods based on deep learning.