Loop closure detection is a crucial module in simultaneous localization and mapping (SLAM), which reduces accumulative error in building environment map. Traditional appearance-based methods for loop closure detection are vulnerable to environmental variations as they mainly rely on hand-crafted features. The convolutional neural networks (ConvNets) can automatically learn feature representation from original image, and it is more robust to illumination changes. However, the ConvNets methods may fail when the viewpoint changes significantly due to it extract global features. In order to solve the problem mentioned above, in this paper, we design an unsupervised network which combines the advantage of the traditional and ConvNets methods, and propose a new module named spatial pyramid pooling based convolution autoencoder (SPP-CAE). We evaluate the performance of the proposed method on several open datasets using precision-recall metric. The results show that our method is feasible for detecting loops and is more robust than state-of-the-art methods.