以人臉照片辨識不良寵物飼主之研究;Identifying bad pet owners from pictures with CNN models

NCUIR > School of Management at National Central University > enterprise resource planning of Optics and Photonics > Electronic Thesis & Dissertation > Item 987654321/88312

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/88312

Title:	以人臉照片辨識不良寵物飼主之研究;Identifying bad pet owners from pictures with CNN models
Authors:	趙千儀;Chao, Chien-Yi
Contributors:	會計研究所企業資源規劃會計碩士在職專班
Keywords:	寵物棄養;虐待動物;不當飼養;飼主;深度學習;人臉辨識;卷積神經網路;Pet abandon;Pet abuse;Improper feeding;Owner of the pet;Deep learning;Convolutional Neural Network;CNN
Date:	2022-06-21
Issue Date:	2022-07-13 22:46:15 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著人口結構和社會經濟型態的改變，越來越多人選擇飼養寵物作為生活伴侶，擁有寵物也已經被證實具有許多的好處。然而，雖然很多人都知道擁有寵物可以帶來許多好處，但寵物遺棄、虐待動物和不當飼養事件至今還是相當普遍。而絕大多數飼主其實就是潛在的棄養者、施虐者，我們無法知道哪一位飼主在未來某一天會棄養，甚至虐待自己的寵物，若是能將潛在不良飼主第一時間排除在外，在第一步就做好預防，進而能有效抑制一連串問題的發生。因此，本研究嘗試透過社群媒體上蒐集黑名單白名單人臉照片，利用深度學習中的卷積神經網路（CNN）演算法進行寵物飼主的好壞辨識。本文的人臉識別CNN 模型結構，由六個卷積層和四個池化層，以及全連接層中的兩個隱藏層組成，我們使用了 Data augmentation（數據增強）來解決數據集不足的問題，使用Batch Normalization（批次正規化）克服模型難以訓練的問題，使用 Dropout 等方法減緩過擬合，並且使用 Adam 優化器和 Softmax 分類器進行人臉識別可以使訓練更穩定、更快收斂，有效提高準確率。透過卷積神經網路很強的特徵提取圖像辨識的方式，可以快速有效地對寵物飼養人進行辨別。實驗結果表明，本研究 CNN 模型在寵物飼主人臉照片上的識別率為 80.09%。;Along with the change in population structure and the social-economic pattern, adopting pets have become popular. However, as raising pets requires long term devotion and commitment, some pet owners may change their minds or may resort to abusing pets to release their emotion. To preventing pets from becoming the victims of bad owners, how to deter the potential abusers to adopt pets has become an important issue. As people with unstable emotion tend to have certain types of facial expression, this study proposed to build up discriminant model for bad owners by analyzing pictures. Even though analyzing facial images with Nero networks have been studied in many areas, to the best of our knowledge, no one has applied related knowledge to this issue. The proposed model utilizes CNN with six Convolution layers, four Pooling layers, and two Fully Connected Layer at the end. As the number of pictures was not sufficient, Data augmentation was utilized to increase the data size. Batch Normalization is utilized to quickly converge the model parameters as the number of data are still relative limited. Dropout and regularization methods are also adopted to relieved the issue of overfitting. Numerous hyper parameters tuning were attempted, and result showed that the accuracy can reach 80.09%.
Appears in Collections:	[enterprise resource planning of Optics and Photonics] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	101	View/Open

社群 sharing

Loading...