Image Classification Based On Capsule Networks

In recent years,Capsule Network(CapsNet)has gradually become a hot research topic in the field of image classification due to its stronger robustness than convolutional neural networks in the study of image classification problems where images are transformed with affine transformations such as shear mapping and rotation.However,there are still many problems in the current capsule networks.For example,the shallow architecture of the capsule network can easily lead to insufficient feature extraction,which in turn leads to low classification accuracy on large-scale data sets containing complex objects.In addition,the activation function used is likely to cause the problem of gradient disappearing,the routing algorithm between capsule layers is too computationally expensive,and the robustness of the model is not strong enough.These are all problems that the capsule network needs to solve urgently.In view of the above problems,this thesis mainly conducts research from the following aspects:(1)Aiming at the shortcomings of the traditional capsule network architecture,this thesis proposes a capsule network called Dense Caps(Dense CapsNet)in order to obtain higher classification accuracy on complex data sets.Dense Caps uses a Dense block as a feature extraction module in front of the capsule layer to extract rich features for the capsule layer.In addition,in order to solve the problem of vanishing gradient caused by the Squashing activation function,this thesis proposes a new activation function namely Caps Re LU.The use of Caps Re LU can not only effectively solve the gradient disappearing problem,but also help solve the parameter initialization problem.(2)Aiming at the huge computational burden of traditional routing algorithms,a new type of capsule network called FR-CapsNet(Fast Routing CapsNet)is proposed.The fast routing algorithm in FR-CapsNet does not require Softmax and for loops to force each low-level capsule to select only one high-level capsule,allowing the information of low-level capsules to be sent to all high-level capsules at the same time.In addition,the proposed routing algorithm uses statistical characteristics like the mean and variance of the predicted values of the low-level capsule to the high-level capsule to cancel the iterative update of the capsule weight in the traditional routing algorithm,which effectively reduces the computational burden.Furthermore,FR-CapsNet also includes an improved dense residual module.Using this module as a feature extractor helps to solve the feature reuse problem in the capsule network.(3)Aiming at the problem that the traditional capsule network is not robust enough to affine transformed images,a novel capsule network RE-CapsNet(Robustness Enhanced CapsNet)is proposed.Unlike traditional capsules that try to learn the invariant spatial relationship between the part and the whole of the object,RE-CapsNet is designed to learn the equivariant spatial relationship between the parts.The model uses an attention mechanism to decompose the object into multiple parts and build primary capsules.Then,by adding a regularization term to the loss function in the training process,the pose of the parts of the object are equivariant to the affine transformation,while ensuring the invariance of the feature descriptors.Based on the above working principle,the RECapsNet model proposed in this thesis achieves better robustness to affine transformed images.In order to verify the effectiveness of the capsule network proposed in this thesis,a comparative experiment with several types of advanced capsule networks was carried out on public data sets.The experimental results show that the capsule network models proposed in this thesis have good performance on image classification tasks and a good practical applicability.