Intelligent Detection And Recognition Of Wild Animals Based On Deep Learning

The rapid development of camera technology and the significant reduction in hardware costs have made the acquisition of photo images less expensive.Camera surveillance equipment is widely used in scientific research,security,transportation,logistics and other fields.In the field of natural ecology,in order to monitor and record wildlife in nature for a long time,study the survival and migration of species,and protect endangered wildlife,various nature ecological research institutions and wildlife protection organizations around the world have densely laid a large number of camera traps in various natural ecological areas,captured wildlife within the monitoring range,and collected a large amount of image data.How to process and analyze these image data quickly and cheaply is an important problem to be solved.In recent years,the deep learning technology based on neural network has developed rapidly,and the computer has become more and more capable of image recognition and analysis.Convolutional neural network algorithm can automatically learn the depth features of images and help computers to realize classification,detection and segmentation.The disclosure of large-scale image datasets in various fields provides abundant resources for the training of neural network algorithms,and the rapid development of computer hardware has greatly improved the execution speed of algorithms.With the emergence and development of computer technology such as deep learning,ecological researchers can automatically perform data cleaning,species identification,counting,target positioning,etc.through algorithms and programs,saving a lot of time and effort.However,at present,the detection and recognition technology of wildlife images still has problems and challenges such as scarcity of image surveillance labels,extremely uneven data distribution,special scenes of infrared cameras,and application of technology.This paper mainly studies the intelligent detection and identification of natural wildlife images based on deep learning.In response to the above problems,this paper proposes solutions and improvement programs.The main contributions are:1)Propose an effective object detection training mechanism.Through object detection of images with classified labels,pseudo-labels are given to the object bounding box,and weak supervision training is realized,and a higher-performance wild animal multi-species detector is obtained,which solves the problem that wild animal image detection labels are scarce and data distribution is uneven.The comparison with the baseline shows that the proposed mechanism can effectively improve the accuracy of object detection of wild animals,especially the identification of rare species with scarce labels,and increase the number of categories of detectable species.2)Propose a domain migration mechanism for infrared images.The Generative Adversarial Nets are used to migrate the wild animal image of the nighttime infrared scene to a color image of the visible scene,and then the original image and the generated image are detected and identified.Experiments on three kinds of wild animals show that the proposed mechanism can effectively improve the recognition accuracy of nighttime infrared images,and further improve the accuracy through the fusion strategy.3)Training and testing models for Asian elephant identification needs in Yunnan and developing application systems.The previously obtained wildlife detection model uses the Asian elephant image from the Xishuangbanna region of Yunnan for transfer learning,and a new detection model suitable for the scenario and task is obtained.Then,based on this network model,an offline detection program and an online call interface are designed and developed,which are provided to organizations in Yunnan to solve their actual needs.