Tensorized Neural Network Based Multi-Label Image Classification

In the real world image recognition tasks,an image usually contains multiple visible objects and abounds with visual information,which results in the need of attaching multiple labels to it.The multi-label image classification algorithms are expected to identify objects included in the image effectively and efficiently,which requires algorithms to extract sufficient features for classification and learn a proper hyper-plane discriminating the inputs.Deep convolutional neural network has abilities of extracting a large number of refined features from images and learning a complex mapping from the high-dimensional feature space to the output space,showing great potential in handling multi-label image classification.However,its high performance is coming with a mass of learnable parameters and high occupation in computer resource,which not only limits developing larger scale networks to tackle classification tasks but also restricts networks' applications in low-end devices.In order to reduce the number of parameters in deep neural networks and improve its performance in multi-label images classification with finite computer resource,this paper proposes Tensor Ring decomposition based neural network tensorizing and tensorized neural network based multi-label image classification algorithm.The main contributions of this paper are as follows:1)The tensorizing neural network method is proposed,which tensorizes fully connected layers using multi-linear transformations between tensor cores as well as linear transformations between tensor cores and inputs.This method helps networks to achieve competitive performance in classification with fewer parameters.2)The multi-label image classification algorithm is presented,which imposes structural constraints on tensor cores to enable efficient feature extraction for specific and overall labels.The theoretical analysis shows that a neural layer can extract features for different labels with varied weights in Tensor Train format in the proposed algorithm,which promotes effective visual information extracting.3)Experiments are carried out on several multi-label image classification datasets,and multiple evaluation measures are employed to analyze experimental results.It shows that the proposed Tensor Ring decomposition based tensorizing neural network method reduces the number of parameters significantly without much accuracies harming,and the multi-label image classification algorithm demonstrates higher performance than recent algorithms.