| Summary: | 碩士 === 國立暨南國際大學 === 電機工程學系 === 107 === In recent years, the application of artificial intelligence (AI) is one of the most highly regarded research topics. The convolutional neural networks (CNNs) have achieved a great development of AI. It consists of multiple convolutional layers followed by pooling layers and fully connected layers. The main purpose of pooling is to reduce the dimensionality of an input image such that the computational complexity can drop efficiently. In this paper, LeNet-5, one of the most classical CNN architectures proposed by LeCun et al. in 1998 [1], has been examined on various datasets in order to experiment and explore the impacts by means of different sorts of pooling schemes in image recognition. For the improvement on both maximum pooling and average pooling [10], an advanced scheme, called adaptive pooling has been proposed. According to the experimental result, adaptive pooling has improved the accuracy on the Adience database by about 1.54%.
Keywords: Deep Learning, Deep Convolutional Neural Network, Pooling.
|
|---|
