Lightweight Human Ear Recognition Based on Attention Mechanism and Feature Fusion

Abstract

With the development of deep learning technology, more and more researchers are interested in ear recognition. Human ear recognition is a biometric identification technology based on human ear feature information and it is often used for authentication and intelligent monitoring field, etc. In order to make ear recognition better applied to practical application, real time and accuracy have always been very important and challenging topics. Therefore, focusing on the problem that the mAP@0.5 value of the YOLOv5s-MG method is lower than that of the YOLOv5s method on the EarVN1.0 human ear dataset with low resolution, small target, rotation, brightness change, earrings, glasses and other occlusion, a lightweight ear recognition method is proposed based on an attention mechanism and feature fusion. This method mainly includes the following several steps: First, the CBAM attention mechanism is added to the connection between the backbone network and the neck network of the lightweight human ear recognition method YOLOv5s-MG, and the YOLOv5s-MG-CBAM human ear recognition network is constructed, which can improve the accuracy of the method. Second, the SPPF layer and cross-regional feature fusion are added to construct the YOLOv5s-MG-CBAM-F human ear recognition method, which further improves the accuracy. Three distinctive human ear datasets, namely, CCU-DE, USTB and EarVN1.0, are used to evaluate the proposed method. Through the experimental comparison of seven methods including YOLOv5s-MG-CBAM-F, YOLOv5s-MG-SE-F, YOLOv5s-MG-CA-F, YOLOv5s-MG-ECA-F, YOLOv5s, YOLOv7 and YOLOv5s-MG on the EarVN1.0 human ear dataset, it is found that the human ear recognition rate of YOLOv5s-MG-CBAM-F method is the highest. The mAP@0.5 value of the proposed YOLOv5s-MG-CBAM-F method on the EarVN1.0 ear dataset is 91.9%, which is 6.4% higher than that of the YOLOv5s-MG method and 3.7% higher than that of the YOLOv5s method. The params, GFLOPS, model size and the inference time per image of YOLOv5s-MG-CBAM-F method on the EarVN1.0 human ear dataset are 5.2 M, 8.3 G, 10.9 MB and 16.4 ms, respectively, which are higher than the same parameters of the YOLOv5s-MG method, but less than the same parameters of YOLOv5s method. The quantitative results show that the proposed method can improve the ear recognition rate while satisfying the real-time performance and it is especially suitable for applications where high ear recognition rates are required.