Research And Implementation Of Bird Fine-grained Image Classification Based On Parent-child Network

With the rapid development of deep learning,fine-grained image classification tasks have received widespread attention.These images are difficult to distinguish due to similar object features,similar poses,and strong background interference.Therefore,correctly classifying such images has become a hot and challenging research topic,In practical production,bird images are difficult to distinguish due to these characteristics.At the same time,as an important part of the ecological world,birds have a significant impact on human production and life.Therefore,using deep neural networks to correctly classify them is of great significance.This paper focuses on studying a classification algorithm for finegrained bird images using the Vision Transformer neural network.The algorithm is based on publicly available bird datasets from the California Institute of Technology and Cornell Lab of Ornithology,as well as a manually constructed airport bird dataset.It is divided into three stages:in the first stage,the Vision Transformer network trains the ancestor network;in the second stage,the ancestor network builds the parent network through knowledge distillation to determine the difficulty of image classification and adopts a "no prediction" strategy for "difficult-to-classify" categories;in the third stage,the parent network builds the child network through knowledge distillation,and a label smoothing strategy is used for "easy-toclassify" categories.The article explores the classification ability of the three-stage network model from both the neural network itself and the data set division,using four meta-learning training methods,Based on the model obtained by the algorithm,a bird image recognition judgment system is implemented and tested in detail for user management,data management,model management,and judgment analysis modules.The algorithm proposed in this paper defines the difficulty of image classification using knowledge distillation based only on image-level labels and handles the results differently to improve the model’s performance.The algorithm achieved good results on three bird image datasets,with classification accuracies of 91.89%,91.34%,and 97.15%,respectively.This method of defining the difficulty of image classification also provides a new research idea for the classification of fine-grained bird images.Furthermore,the test results of the system show that it can store bird image information well,manage the models trained by the algorithm,and ensure accurate classification of bird images by users.It can be well applied to bird identification and conservation.