Wild Animals Object Detection Based On Deep Learning

Wildlife diversity monitoring is a prerequisite,foundation,and key component of wildlife research,protection,and management.Most nature reserves install camera traps to obtain images,which are accumulated in tens of thousands of units.These image data contain a large amount of key information about wildlife,such as species composition,population size,population dynamics,etc.These data mainly rely on artificial visual screening and recognition,which is far behind the speed of image data accumulation,and seriously restricting the effective application of data in research,protection,and management work.Early traditional object detection methods used supervised machine learning to reduce time and labor costs to some extent,but these methods required manual setting of extracted features,resulting in weak model generalization ability.Later,deep learning was widely applied to object detection,but the model required a large amount of dataset for training.Specifically,for wildlife object detection tasks,obtaining wildlife images is difficult,resulting in insufficient or missing data for certain species,making it impossible to construct a balanced dataset,and sometimes resulting in low accuracy of trained model.In order to solve the problem of small dataset size or imbalanced data in the dataset when detecting wildlife object,this paper proposes a few-shot wild animals object detection method based on fine-tuning,which is divided into two stages.The first stage is to identify and locate wild animal object from the image,and the second stage approximates the contour of the first stage detection animal object bounding box to achieve contour based wild animal segmentation.The main research content of this paper is as follows:(1)Create a wildlife dataset suitable for the research methods in this paper.Due to the limited number of animal species in existing large-scale public datasets such as Image Net,PASCAL VOC,COCO,and other comprehensive datasets,and the fact that most of them are pet or livestock animals such as cats,dogs,birds,etc.,it is necessary to construct wildlife images that meet the requirements of this paper.The dataset constructed in this article is partly based on the required wild animal species images from the large public dataset mentioned above,while the other part extracts frame images from documentary videos to supplement the dataset appropriately.Annotate the dataset and create a wildlife dataset suitable for deep learning object detection.(2)Implement a few-shot wild animals object detection algorithm based on fine-tuning.A few-shot wild animals object detection algorithm is proposed.The algorithm uses the Faster R-CNN network architecture,and the Res Net-101 deep learning backbone network is used to extract the characteristics of wildlife species;At the same time,the combination of feature pyramid(FPN)and backbone network can fuse multi-scale wildlife species characteristics;The training method adopts a phased training method: in the first stage,the dataset is used to train the network and save the weight parameters.In the second stage,a small and balanced sub dataset containing new classes is constructed,and the weight parameters of the fixed feature extraction network are fixed.The position regression and classification prediction network is retrained to obtain new weight parameters,ultimately achieving accurate recognition of wildlife.(3)Research on contour approximation strategies for wildlife object detection based on deep learning.Taking the wild animal bounding box detected and output in research content(2)as input,first take the midpoint of the four edges of the bounding box and connect it into a diamond bounding box.After passing through the convolutional network for the first time,generate the pole coordinates for predicting the positions of the upper,lower,left,and right sides;Then each side extends a certain length to both sides with the pole coordinate as the center and forms a octagon bounding box,which is sent into the convolution network for the second time to approximate the contour of the octagon bounding box.By iteratively updating the bounding box multiple times,gradually approaching the contours of wildlife,and finally achieving accurate extraction of wildlife contours.(4)Based on the overall composition of the system and equipment deployment,a wild animal object detection software platform is developed to automatically detect massive wildlife monitoring images collected by camera traps,detect and extract wildlife contours,and output and display them through a visual interface.