卷積神經網路於影像辨識之應用

Applications of convolutional neural networks on image recognition

郭衿懿 ； 壽大衛

《NCS 2019 全國計算機會議》 2019 (2019/11) Pp. 677-681

https://doi.org/10.6927/NCS.201911.0132

摘要

本研究提出透過深度學習之影像辨識技術，利用卷積神經網路（CNN）所具有很強的特徵提取功能，透過濾波器（filters）一層一層從輪廓、邊緣線條到局部特徵自動進行提取，再以完全連接（fully connected）方式將特徵資料輸入至神經網路進行訓練，由於對於影像與語音辨識處理效果很好，易於找出重複出現的特徵，有利於差異辨識以進行分類。使用卷積神經網路，透過參數與神經網路架構之調整，以貓、狗為研究案例，應用於流浪動物通報多元管道。流浪動物主要是貓與狗，因為貓與狗兩種動物之體型、大小、毛色、外觀等特徵相類似，利用卷積神經網路提取個別特徵再加以訓練之特性，可提高辨識度，並有助於協助簡化流浪動物通報方法。實驗結果顯示，本研究提出之方法，辨識率最高可達98%，利於流浪動物之通報。

關鍵字

影像辨識 ； 深度學習 ； 卷積神經網路

並列摘要

This study proposes the use of image recognition technology through deep learning. By utilizing the strong feature extraction function in the convolutional neural network (CNN), we are able to extract automatically contours, edge lines, and local features layer by layer through filters. We then fully connect the data and input them to the neural network and train the CNN. Due to the effectiveness of image and voice recognition, it is easy to find recurring features and classify the differences. By adjusting of parameters of CNN, we're able to apply the technology in reporting stray animals in different channels. Stray animals are mainly cats and dogs. Because cats and dogs are similar in body size, size, coat color, and appearance, the characteristics of using individual convolutional neural networks to extract individual features and training can improve recognition. Therefore, it helps to simplify the reporting mechanisms of stray animals. The experimental results show that the proposed method in this study increases the recognition rate of up to 98%, which is beneficial to the reporting of stray animals.

並列關鍵字

image recognition ； deep learning ； convolutional neural networks (CNN, or ConvNet)