Rotational Objects Recognition and Angle Estimation via Kernel-Mapping CNN

Yuanyuan Zhou,Chen Wang,Xiaoling Zhang,Xiaqing Yang,Jun Shi,Shunjun Wei

doi:10.1109/access.2019.2933673

Abstract

Convolutional neural network (CNN) has become the mainstream method in the field of image recognition for its excellent ability to feature extraction. Most of the CNNs increase the classification accuracy for the rotational objects by imposing the network with rotation invariance or equivariance property, which causes the loss of the targets orientation information. This paper attempts to achieve objects recognition and angle or orientation estimation simultaneously without additional network training. To this end, we propose the matching criterion and the kernel-mapping convolutional neural network (KM-CNN). It has been shown that when the kernel satisfies the matching criterion, the output remains the same. Based on this study, we apply rotation transformation to the KM-CNN. Besides, the KM-CNN with the rotation by shifting pixel method and octagonal convolutional kernels can solve the mismatching problem caused by the rotations. The KM-CNN with the kernel sharing central weights gives the near state-of-art results in target recognition and angle estimation on benchmark datasets MNIST, GTSRB and Caltech-256.

Highlights

Convolutional neural networks (CNNs) have become one of the hottest spots in the field of machine learning
In this paper, we researched the spatial characteristics of convolutional neural networks
A kernel-mapping CNN implemented by rotating the convolutional kernel is proposed, which can recognize and estimate the rotation angles

Summary

Introduction

Convolutional neural networks (CNNs) have become one of the hottest spots in the field of machine learning. It is for the reason that CNNs, with wide applications, have achieved many stateof-the-art results in various practical tasks, such as image classification [1]–[3], object detection [4], [5] and semantic segmentation [6]. Comparing with the traditional machine learning methods, the significant results of CNNs owe to the embedded convolutional layers. CNN performs well in targets recognition and is compatible with targets translation and scaling to a certain extent, it is sensitive to targets rotation.

Methods

Findings

Discussion

Conclusion