Research And Application Of Image Knowledge Extraction Technology Based On Deep Learning

Image knowledge extraction technology is an important research direction in computer graphics.It is a technique for extracting knowledge from various aspects such as features and categories of images.The main research direction of this thesis is the knowledge extraction technology of fine-grained images classification.Due to the current deep learning field,fine-grained classificaition of images which in same category remains a challenging task.The experimental environment mainly focuses on two different environments: strong supervision and weak supervision,the use of convolutional neural networks can further improve the experimental accuracy.The point of this research is on image knowledge extraction and image fine-grained classification in these two environments.In this thesis,the fine-grained classification under these two experimental environments has been explored,and some progress has been made in accuracy and performance.At the same time,this thesis has also been implemented in the actual project,using the above methods to extract the knowledge of fine-grained classification image from a group of similar plant leaf images.The main work of this thesis is as follows:1.Introduced the convolutional neural network in deep learning,mainly adopting the experimental environment under strong supervision,and using certain annotation work on the image,and the main R-CNN neural network is improved according to the scene.2.Two different methods are chosen for the experiment under weak supervision.The convolutional neural network was designed,trained,and applied in the fine- grained classification experiment of leaf image dataset.The experimental results are recorded and the advantages are analyzed.Some optimization schemes are put forward for the related problems.3.After completing the experiment in the weak supervision environment,a complete high-performance concurrent server is designed and deployed.So the whole system can complete a series of network transmission image recognition services.After completing the main work,we combined with the current various methods for comparison,and also combined with the experimental results in different environments in this thesis for analysis and comparison.The experimental results show that the image knowledge extraction algorithm designed in this thesis can effectively complete the task of image knowledge extraction and has a good performance on the dataset.