EIQA: ear image quality assessment using deep convolutional neural network

It is well established that biometric image quality plays a pivotal role in the overall performance of biometric system. In poor-quality images, detecting true features is complicated and is the main reason for matching errors. The quality of ear images in unconstrained environment conditions poses many challenges to the ear biometric system. To tackle the aforementioned issues, in this work, we have proposed a non-reference based ear image quality assessment (EIQA) network and explored its usage in ear recognition system. The quality assessment of ear images is a challenging task as there doesn’t exist any database that provides quality score of ear images. The quality of ear images in the unconstrained environment is mainly affected by six different distortions such as huge rotation, blurriness, white noise, high occlusion, poor contrast, and low resolution. These distortions are incorporated at five different levels to prepare different bins of image quality. The network is trained to identify the type of distortion and classify the image into five different categories viz. excellent, good, fair, poor, and bad. The proposed method gives an overall classification accuracy of 97.46% on images that include all types of distortions. Afterwards, the experiments are performed for ear recognition, and it has been identified the quality of the ear images highly affects the system’s performance. An overall Correct recognition rate (CRR) is improved from 24.90% to 72.45% by discarding the images with an inferior quality. Thorough experiments are performed on IITK ear recognition database. To the best of our knowledge, this is the our first attempt made to develop self semi supervised network training strategy to identify, quantify and estimate the quality of ear images.This work aims to improve the performance of the ear recognition framework by accurately assessing the quality of images and discarding the low quality images. The results demonstrate the effectiveness of the model and a viable option to be utilised in the improvement of ear recognition methods.