THE EFFICIENCY OF THE TENSORFLOW MODELS IN THE APPLICATION TO THE TASK OF DETECTION OF EYES IN THE PHOTO
S. O. Burdukowsky
Novosibirsk State University of Economics and Management, Novosibirsk, Russian Federation
Keywords: object detection, machine learning, eyes, artificial neural network, classification loss, localization loss, detection accuracy
Abstract
In this article, the author compares the effectiveness of TensorFlow detection models in solving the problem of detecting areas with eyes in human face photo. Experiments of two types were carried out: additional training of a pretrained detection model and training of the model from scratch. Face images from Flickr-Faces-HQ Dataset were used to form training and evaluation samples. The article describes the training parameters, shows classification and localization loss graphs, assesses the accuracy of the trained models, and also demonstrates the operation of the “SSD MobileNet V2 FPNLite 320×320” detection model trained from scratch, which received the highest accuracy scores after additional training and training from scratch. For programs with a requirement for IoU of detected objects greater than 0.5, the accuracy of the model is 99.9 %. The results of the experiments can be applied in various researches, that use the TensorFlow platform to detect objects in images, and only one class of objects is detected.
|
