초록 close

여러 인공 지능 분야에서도 이미지 분류기는 가장 근본이 되는 분야이며 또한 가장 널리 사용되는 분야이다. 이 논문은 이미지 분류기의 정밀도 성능을 올리기 위한 새로운 방법을 제안한다. 사물을 인식할 때, 사물의 형상을 보고 판단한다는 직관으로부터 새로운 신경망 구조를 구현하였고 은닉층의 특징 맵을 상대 비교함으로써, 필현적으로 성능향상이 이루어질 수 있음을 보였다. CIFAR-10 실험 세트를 이용하여 약 3% 정도의 성능향상을 보여, 제안하는 방법이 실제 유효함을 보였다.


Among the several deep learning research areas, image classification is a fundamental area and has been widely applied to many practical applications. Image classification is to determine which category the given image belongs to. Since the image classification is a typical supervised learning, test images and the corresponding answers, i.e., labels are also given. And, with the given test images and labels, a neural network is trained to minimize a loss function defined by error between the label and the inference result. Therefore, as the loss decreases, the inference accuracy increases. As a result, the accuracy is a criteria of the performance of the neural network in image classification. In this paper, a new method for high accuracy image classification is suggested. Authors thinks that recognizing things mean seeing the shape of the things. With the intuition, in the proposed method, additional edge information is applied to the image and around 3% accuracy improvement is achieved in the experiment. In order to clarify the improvement, compared feature maps in the hidden layers are visualized and analyzed. And, it is also confirmed that the feature map of the proposed method is more clear and sharp than that of the conventional one. Another merit of the proposed method is that this method can easily improve the accuracy of the conventionally existing neural networks through the transfer learning because the proposed method just modifies the first layer of the conventional neural networks.