Capsule Network Approach for Image Classification

Numerous deep learning (DL) models based on convolutional neural network (CNN) have been developed for image classification. Though the CNN models attained outstanding performance, its architecture contains certain limitations such as loss of information in the pooling layer and no spatial relationship among the features of an object. The development of capsule network (CapsNet) has overcome the drawbacks of CNN. The CapsNet works on a set of capsules, and each capsule comprises a set of features. The CapsNet consists of a convolution layer, a primary capsule layer, image caps and a fully connected (FC) layer. In this paper, a variant of CapsNet approach is proposed by tuning the parameters of the original CapsNet and replacing ReLU with AReLU activation function in each FC layer. The classification performance of the proposed approach is evaluated on the Plant Village (PV) and Tomato datasets. 80% of the datasets were trained and 20% were tested. The proposed approach considering filter size of 5 × 5 in the feature extraction layer with AReLU activation function in each FC layer delivers 94.02% and 94.06% classification accuracies on PV and Tomato datasets respectively. Compared to other modern approaches, the proposed approach delivered better results.