CNN Based Wildlife Recognition with Super-Pixel Segmentation for Ecological Surveillance

Recent years, the convolutional neural network have shown to provide excellent results on recognition in different competitions. However, challenges in specific missions still exist. The cluttered backgrounds and rich feature changes of wild environment bring great challenges to the problem of species recognition of wild animals. To address these problems, this paper proposes a novel and effective combination to learn a CNN model. This is achieved by apply simple linear iterative clustering (SLIC)super-pixel segmentation method to unified data dimension during the process of making raw image data (captured by camera-traps)into a dataset. In short, the super-pixel-divided images provides the input of the convolutional neural network. In order to verify the application, we conducted a comprehensive performance comparisons between our SLIC-dataset and generally used Resize-dataset over CNN networks. Results proved that our proposed method performs exceptionally well in low-resolution data when it is crucial to take full advantage of the edge information of original images. In addition, we collected and annotated a standard camera-trap dataset of 14 common wildlife species in China, which contains 16,480 training images and 4,120 testing images.