Hide/Show Apps

A Multi-camera system for automation of mouse grimace scaling using convolutional neural networks

Ağca, Ahmet
Over the past decade, convolutional neural networks (CNNs) have gained great progress on the area of computer vision. Many problems related to automation of image recognition or classification are now possible to be solved using CNN with an accuracy much more than a human can achieve. One of these problems is the automation of Mouse Grimace Scaling (MGS). It is such a time consuming and error-prone task even for an expert to classify the pain levels of a mouse for lots of images captured from videos. For this reason, it is essential to incorporate the benefits of popular machine learning algorithms into this research area. The purpose of this thesis is to achieve significant results for practical implementations along with improving the methodology for automation of MGS. In this thesis, a complete set of methodology starting from the mouse monitoring setup to building the neural network model for automation of MGS was studied and the results were compared with that of previous works. For detecting the mouse in video frames, the previously developed tracker algorithms and detection networks were used without change. The evaluation was performed by means of both classification and regression problem and transfer learning was adopted for the basis of the study. For regression, MAE of 0.226 was achieved for one cross-validation (CV) balanced set, and 0.26 was achieved for overall balanced sets (score range, 0 to 2). For binary-classification, 91.10% accuracy was achieved for one CV balanced set, while 82.45% was achieved for overall balanced sets