應用卷積類神經網路於生態觀測影片之動物識別研究－以蓮華池試驗林為例

本研究主要目的是使用卷積類神經網路(convolutional neural network, CNN)進行動物影像之自動化偵測與辨識技術。本研究方法使用了YOLO(you only look once)的卷積類神經網路架構，先針對以動物科名將影像中的移動物件做偵測與訓練，然後再對此物件訓練結果進行框選與切割，由於YOLO偵測特徵太相似的動物效果不好，再經由GoogleNet Inception進行以動物學名細部辨識遷移學習辨識訓練，以便提升動物學名細部的辨識分類的準確度。為了驗證本文方法有效性，本研究做了動物偵測與辨識的實驗，實驗影像取自蓮華池研究中心動物生態觀測影片資料，每部影像長度為30秒影像資料集分為二個部分，第一部份為動物科名偵測的資料集訓練，其資料集有700張影像，140張測試影像，第二部份為動物學名辨識的資料集訓練，其資料集有360張影像，90張測試影像。經過實驗證明，本研究的動物偵測結果，其偵測率為85.71%。至於動物細部辨識實驗結果，其辨識率高達88.88%。最後也做了動物辨識處理時間的計算，本研究方法在1秒內可以辨識10張的影像，確實節省了人工龐大的辨識時間，也證明了本文方法可以有效快速的處理大量影像。從偵測與辨識實驗結果，顯示本文方法可以有效且快速的進行動物偵測與辨識，期望未來可以將本研究成果運用在自動相機動物影像的偵測與辨識上。

關鍵字

卷積類神經網路；野生動物；試驗森林；偵測；識別

並列摘要

The main purpose of this study is to use the convolutional neural network (CNN) for automated detection and identification of animal images. This research method uses the YOLO (you only look once) convolutional neural network architecture, which first detects and trains moving objects in the image under the name of the animal, and then frames and cuts the training results of the object. Because YOLO detects when the characteristics of the animals are too similar, it then uses GoogleNet Inception to carry out the identification training of the animal name, in order to improve the accuracy of the identification classification of the animal name details. In order to verify the effectiveness of the method, the experiment was performed on animal detection and identification. The experimental images were taken from the animal ecological observation film data of the Lienhuachih Research Center. Each image length was 30 seconds. The image data set was divided into two parts. In the part on the training for the collection of animal names, the data set has 700 images and 140 test images; the second part is the data set training for animal name identification. The data set has 360 images, and 90 sheets of test images. The experimental results show that the detection rate of the animal in this study is 85.71%. As for the animal detail identification experiment results, the recognition rate is as high as 88.88%. Finally, the processing time spent on the calculation of animal identification was also done. This research method can identify 10 images per second, which saves artificial large recognition time, and proves that the method can effectively process a large number of images efficiently and quickly. The results of the detection and identification experiments show that the method can effectively and quickly perform animal detection and identification. It is expected that the research results can be applied to the detection and identification of automatic camera animal images in the future.

並列關鍵字

convolutional neural network ； wild animals ； experimental forests ； detection ； identification

參考文獻

[1] 行政院農委會林業試驗所，蓮華池研究中心介紹，取至：https://www.tfri.gov.tw/main/page_view.aspx?siteid=&ver=&usid=&mnuid=5347&modid=1339&mode=&cid=1334&noframe=#page_con，資料取得日期：民108年7月19日。