CapsNet based on Encoder and Decoder for Object Detection

摘要

The recently proposed capsule network (CapsNet) can learn the hierarchy relationships of entity features and realize the equivariance to affine transformations, which makes the capsule architecture more promising for object detection. In this paper, based on capsule architecture, we create the CapsNet-V1 models for object detection. The proposed CapsNetV1 mainly consists of the classification net as encoder to extract multi-class information and the reconstruction net as decoder to obtain masks with multi-object position information. In the experiments, based on the randomly expanded MNIST dataset, we simultaneously evaluate the multi-object classification and reconstruction abilities of the proposed CapsNet. The results indicate that our capsule models can reconstruct the object masks with accurate location information at correct labels, which exactly demonstrates the feasibility of using capsule networks for object detection. Further, our CapsNet can be widely applied to the multi-object detection with simple backgrounds in the industrial production lines.