以深度學習為基礎之野生動物辨識系統
No Thumbnail Available
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
近年來人們對於動物保育的意識抬頭,越來越多人加入保護野生動物的行列,對於野生動物生態系統進行觀察可以為人們提供保育的方針。目前主要觀察的方法是使用陷阱相機,因為其能夠長時間運行且不對野生動物造成影響。但由於陷阱相機擷取影像數量太過龐大,分析影像變成耗費大量的時間與勞力的枯燥工作。故本研究擬開發一套以深度學習為基礎之野生動物辨識系統,能夠自動辨識與計算影像中動物種類與數量,以期達到輔助分析之結果。野生動物辨識系統使用陷阱相機所攝影像來進行動物物種辨識,透過物件偵測類神經網路模型來辨識影像中野生動物的種類以及數量。本研究使用RefineDet物件偵測模型的改良版來進行動物物種辨識,將野生動物影像輸入本系統後,經由物件偵測類神經網路模型進行動物物種辨識以及數量統計。在改良RefineDet原型架構方面,本研究引入彈性非極值抑制演算法、受納域模塊等改良來提升RefineDet模型對野生動物的辨識能力;另外加入批量正規化技術來加速模型的訓練過程,提升RefineDet的整體效能。本研究使用賽倫蓋蒂陷阱相機資料集(Snapshot Serengeti dataset)[2]所蒐集的影像進行訓練及測試,辨識的野生動物種類共有11種,分別是水牛(Buffalo)、非洲象(African Elephant)、葛氏瞪羚(Grant's Gazelle)、湯氏瞪羚(Thomson's Gazelle)、長頸鹿(Giraffe)、黑斑羚(Impala)、灰頸鷺鴇(Kori Bustard)、獅(Lion)、牛羚(Wildebeest)、斑馬(Zebra)等。實驗結果為本野生動物辨識系統VOC mAP為83.29%,顯示出本研究所提出之野生動物辨識系統確實能夠準確偵測出影像中野生動物物種及數量。
In recent years, people’s awareness of animal conservation has risen, and more and more people have joined to protecting wildlife. Observing the wildlife ecosystem can provide people with guidelines for conservation. At present, the main observation method is to use a camera trap, because it can service for a long time and does not affect wildlife. However, because the number of images captured by the camera trap are too large, analyzing the images becomes a boring task that consumes a lot of time and labor. Therefore, this research intends to develop a wildlife recongnition system based on deep learning, which can automatically identify and calculate the types and numbers of animals in the images, in order to achieve the results of auxiliary analysis.The wildlife recognition system uses the image captured by the camera trap to recognize the species, and uses the object detection neural network model to recognize the type and number of wildlifes in the image. This research uses an improved version of the RefineDet object detection model for species recognition. After inputting wildlife images into the system, the object detection neural network model is used for species recognition and quantity statistics. In terms of improving the prototype architecture of RefineDet, this research introduces improvements such as soft non-maximum suppression algorithms and receptive field block to improve the RefineDet model’s ability to recognize wildlifes; in addition, batch normalization technology is added to accelerate the model training step and improve RefineDet’s performance.In this study, images collected by the Snapshot Serengeti dataset [2] were used for training and testing. A total of 11 species of wild animals were identified, namely Buffalo and African Elephant, Grant's Gazelle, Thomson's Gazelle, Giraffe, Impala, Kori Bustard, Lion, Wildebeest, Zebra, etc. The experimental result is that the VOC mAP of the wild animal identification system is 83.29%, which shows that the wildlife recognition system can accurately detect the species and quantity of wildlifes in the image.
In recent years, people’s awareness of animal conservation has risen, and more and more people have joined to protecting wildlife. Observing the wildlife ecosystem can provide people with guidelines for conservation. At present, the main observation method is to use a camera trap, because it can service for a long time and does not affect wildlife. However, because the number of images captured by the camera trap are too large, analyzing the images becomes a boring task that consumes a lot of time and labor. Therefore, this research intends to develop a wildlife recongnition system based on deep learning, which can automatically identify and calculate the types and numbers of animals in the images, in order to achieve the results of auxiliary analysis.The wildlife recognition system uses the image captured by the camera trap to recognize the species, and uses the object detection neural network model to recognize the type and number of wildlifes in the image. This research uses an improved version of the RefineDet object detection model for species recognition. After inputting wildlife images into the system, the object detection neural network model is used for species recognition and quantity statistics. In terms of improving the prototype architecture of RefineDet, this research introduces improvements such as soft non-maximum suppression algorithms and receptive field block to improve the RefineDet model’s ability to recognize wildlifes; in addition, batch normalization technology is added to accelerate the model training step and improve RefineDet’s performance.In this study, images collected by the Snapshot Serengeti dataset [2] were used for training and testing. A total of 11 species of wild animals were identified, namely Buffalo and African Elephant, Grant's Gazelle, Thomson's Gazelle, Giraffe, Impala, Kori Bustard, Lion, Wildebeest, Zebra, etc. The experimental result is that the VOC mAP of the wild animal identification system is 83.29%, which shows that the wildlife recognition system can accurately detect the species and quantity of wildlifes in the image.
Description
Keywords
陷阱相機, 物種辨識, 物件辨識類神經網路, 深度學習, 野生動物物種辨識, camera trap, species recognition, object detection neural network, deep learning, wildlife recognition