Abstract - In recent years, biometric recognition based systems have become widespread. One of these is wrist-based recognition systems. In this study, wrist print based recognition system was developed by using near infrared (NIR) camera. Totally 220 NIR camera images taken from 10 for each both hands of 11 people. The obtained data set is allocated 70%(153 images) for training and 30%(67 images) for testing. The wrist regions are labeled on the training set images. The labeled data was trained with YOLOV2 architecture supported by ResNet50 one of the deep neural network models. The trained model was tested with the remaining 30% of the data set. In the test process, the wrist region was determined in the NIR images with the trained model. As a results of the study, it was seen that the wrist regions were correctly detected in all test images and the mean value of obtained similarity rates was %95.26. Therefore, it can be said that the deep learning architectures ResNet and YOLO are effective in the segmentation of the wrist region. Keywords - Wrist Print Recognition, Deep Neural Networks, Near-Infrared Camera, YOLO I. INTRODUCTION IOMETRIC recognition based systems are known as most secure systems. What makes biometric systems important in this way is that the security parameter they use is personal and there is no risk of theft and cannot be copied. Although there are many biometric recognition methods, fingerprint recognition, vein recognition and face recognition are mostly used today. Vein recognition is a high-security biometric recognition approach based on human vascular structure. In this biometric identification approach, finger veins, hand veins, palm veins and wrist veins are used as biometric parameters. The identification systems which use wrist vein structure as biometric input are called wrist print recognition. In wrist print recognition system, the human wrist image is captured by using near-infrared (NIR) camera and illumination and then the wrist region is segmented for determining the wrist vein information. Then, the wrist print is used for identification. In this study, the wrist vein region was detected and marked from the hand and wrist image taken using a near infrared camera. In this context, a software based on deep learning has been developed including ResNet50 and YOLO algorithms. The performance of the system was evaluated comparatively by testing the segmentation process on 220 wrist images taken from 11 people. II. DEEP NEURAL NETWORKS (DNN) Deep neural networks (DNN) can recognize objects without being affected by different properties such as different positions, directions and camera angles and environmental factors such as lighting. A deep learning algorithms are trained on tagged images. An architecture of DNN is shown in Figure 1. Figure 1: A Deep neural networks architecture. The input usually consists of images or signals. In the convolution layers, the filtering process is applied to the 3- dimensional matrices in the previous layer. The number of filters used constitutes the depth (the size of the 3rd dimension) of the convolution layer. In the pooling layer, size reduction is applied. In the section called fully connected layer, classical artificial neural network operations are performed. The output can be defined as a vector with the length of the defined class. The number of layers in the architecture and the filter size and number can be changed by the user to suit the application. In addition, the performance may vary depending on the number of images and iterations to be used for training. If the number of training iterations is too high, both the training time increases and the model moves towards memorization. As the rate of misclassification in data that the model has not seen before will increase as a result of memorization, accuracy decreases. In the study, it was tried to determine the wrist region from the images taken by using the ResNet50 and YOLO architectures. A. ResNet50 The Residual Network (also known as ResNet) uses redundant blocks with multiple layers to reduce training error. Typical ResNet models are implemented with double- or triple- layer skips that contain nonlinearities (ReLU) and batch normalization in between. During training, the weights adapt to mute the upstream layer, and amplify the previously-skipped layer. In the context of residual neural networks, a non-residual network may be described as a plain network [1, 2]. Wrist Print Region Segmentation Based on Deep Neural Networks H. E. KOÇER 1 and K. K. ÇEVİK 2 1 Selcuk University, Konya/Turkey, ekocer@selcuk.edu.tr 2 Akdeniz University, Antalya/Turkey, kcevik@akdeniz.edu.tr B International Conference on Engineering Technologies (ICENTE'20) Konya, Turkey, November 19-21, 2020 27 ___________________________________________________________________________________________________________ E-ISBN: 978-625-44427-4-2