Estimation of Non-trained Category in Image Classification Model based on Deep-Learning
SangYub Ji, Hoyeon Kim, Jae-Soo Cho · Journal of Institute of Control Robotics and Systems · 2018
In this paper, we propose an effective estimation method of non-trained categories in an image-classification model based on deep learning. Non-trained categories can be estimated by using confidence calibration and linear interpolation in the image-classification model. Usual image classifiers that use CNNs (convolutional neural networks) estimate the learned image category with the highest degree of similarity in the previously trained image categories. However, the output results of conventional image classifiers that use CNNs are not probability values of the learned image categories. The image-classifier output values are calibrated to the probability value by using confidence calibration instead of only the maximum category value. Then, we can obtain continuous probability estimates through linear interpolation at discrete output values. The confidence calibration and linear interpolation of discrete output values are post-processing steps that produce linearly calibrated probabilities. The proposed method is applied to estimate the rotation angle of a cup-ramen container, demonstrating the effectiveness of the proposed method and the estimation accuracy of the continuous rotation angle.