Research And Implementation Of Panda Action Recognition Based On Deep Learning

In recent years,with the increasing demand of action recognition technology in practical applications,this technology has become a research hotspot in the field of computer vision.By identifying and classifying the panda material,it can provide a large number of rich and real panda resources to the panda culture creators,so that the panda culture creators can easily extract the required images from the rich resource library.This can greatly facilitate the related cultural creation,improve the efficiency of panda cultural creation,and at the same time reduce the creation cost,which has very important economic and cultural values.For the current research on panda action recognition,there are the following problems:(1)There is currently no standard panda data set.(2)due to factors such as panda shape and background It is difficult to classify the panda action sample set,which makes panda action recognition relatively difficult.(3)The number of samples in the Panda action dataset is relatively small,and there is a problem of low recognition rate of small category samples caused by data imbalance.(4)There is no platform that can actually apply the Panda action recognition algorithm.In view of the above problems,the research contents of this thesis mainly include the following three points.First,this thesis proposes a panda motion detection algorithm under data imbalance and a Transformer-based panda action recognition algorithm.The two algorithms perform action recognition on two different types of data sets: image and video.Among them,the Panda motion detection algorithm under data imbalance improves the SSD model by dividing the training set to improve the recognition rate of small categories in the Panda image data set.The Panda action recognition algorithm based on the Transformer model,based on the Transformer model,adds an optical flow operator to extract motion features between video frames,and improves the recognition rate of the algorithm by extracting effective optical flow features.At the same time,in order to evaluate the algorithm more accurately,we have made datasets of panda pictures and videos respectively,and performed experiments on the two algorithms proposed in this thesis and their benchmark algorithms on the dataset.Experiments show that both algorithms have improved the recognition effect of the original model.Finally,the Panda Resource Library platform was implemented.The platform provides users with a large number of panda material resources,and provides users with a way to upload panda-related pictures and video resources to realize the sharing of panda resources.The platform also uses the algorithm proposed in this article to classify panda materials,and users can accurately retrieve different classified picture and video materials in the database.