Multi-label Learning Based On Convolution Neural Network

Compared with single-tag learning in traditional machine learning,multi-tag learning aims to capture the multi-semantic information contained in data,which has become a research hotspot in the field of machine learning in recent years.Aiming at the problem of multi-label in image classification,this paper constructs an image feature learning model and its application framework,which provides a new solution for multi-label image learning.The main research results are as follows:(1)On the basis of studying the food web stability by dynamics,changing the parameters of dynamic model,exploring the stability by analyzing the dynamic complexity of food web,discovering the extinction relationship,survival dependence relationship,couple which exhibits close relationship and important energy flow path among species in the food web in Somme Bay.(2)According to the newly generated network model,the objective function of network optimization is given in this paper.In order to generate a consistent and balanced binary code,we add two constraints to the objective function,and set the model to learn the image feature binary code by minimizing the classification error defined by the objective function and other ideal binary code attributes.(3)Since model learning is carried out in a point way,it means that it can be extended in multiple data sets.In the single labeled data set,after several depth model replacement,binary code length adjustment and constraint parameter optimization experiments,the model framework can complete the feature extraction from image data to labeled data without losing the performance of the original classification network.(4)Experiments were designed to verify whether the feature binary code learned from the image data contained semantics,and image retrieval was used to verify whether the binary code had category consistency.Then it combines with the existing multi-tag algorithm to realize multi-tag classification.On different data sets of the experiment proved that the presented framework of network structure in the process of learning characteristics of the binary code can retain semantic information,can be generated after conditions has the characteristics of excellent properties of binary code,and based on the characteristics of the binary code can complete image classification and image retrieval task,and good performance on multiple data sets.