- 學位論文
加權自編碼器特徵擷取在影像分類上的應用
Weighted Autoencoder Features Extraction for Image Classification
摘要
深度學習是人工智慧中一個非常實用的技術，而且已經有很多成功的應用，所以現在是一個熱門的議題。利用這個技術去進行一些傳統上無法以簡單的數學方法，或是傳統的機械學習的技術的應用上，往往可以得到令人驚艷的成果。 本論文嘗試用深度學習的方式去探索影像分類的問題。影像分類是基於從影像上擷取特徵，透過不同影像的類別中的特徵來分類影像。然而影像的內容相當複雜，目前仍無法準確定義影像的特徵，因而造成影像的特徵無法透過簡單的方式被擷取出來，因此影像分類仍然是一個急需解決的難題。 在本研究中，將利用擁有深度學習為基礎的自編碼器來提取影像特徵，其中將提出加權損失函數來加強影像中特定區域的特徵，從人類的角度來看使影像分類能更加一致，模擬測試得結果達到令人滿意的表現。
並列摘要
Deep learning is a very applicable technology in artificial intelligence, and there have been many successful applications, so it is now a very popular issue. By deep learning, the applications, which are impossible modeled by simple mathematical method or traditional machine learning, usually achieve desired results. In this work, we try to investigate image classification using deep learning model. Image classification is based on the features that extracted form images. Using the characteristics in each image category, the image is classified accordingly. However, due to the complexity of the images, there is no precise definition for image features. Therefore, the features cannot be extracted in a simple manner. Thus, the image classification is still a difficult problem that needs to address. In this study, we proposed a deep learning based autoencoder for image feature extraction. We developed a new weighted loss function for emphases the image feature in region of interesting, and then the classification can be more consistent for human perception. Simulation results achieves satisfying performance.