Research On Image Retrieval Method Based On Convolution Neural Network And PCA Dimension Reduction

As an important medium for human understanding of the world,vision vividly shows the colorful world for human.Because today's information volume is exponentially displayed in front of people,take image information as an example,it's becoming more and more difficult to find images that are true to mind from big data.Although the traditional technology content based retrieval,for example,based on color,texture,space shape,etc,solved the shortcomings of text based detection,such as heavy workload and subjective factors.But it's understanding of the semantic features of images is not very good.Therefore,in today's increasingly expanding data environment,efficient image retrieval has become one of our goals.In this paper,a new image retrieval method based on convolution neural network and PCA dimensionality reduction is proposed to solve the problem of weak expression of visual features and the high complexity of feature data.First,we extract the semantic features after migration learning convolution neural network and construct the feature library from the image database.We use Inception-v3 model to do migration learning on Corel and Caltech-256 data sets,and get the convolution neural network after migration learning.Secondly,we use PCA to reduce the dimension of the features in the image feature library extracted by the convolution neural network after migration learning.Dimensionality reduction to 64,96,128,and 256 dimensions,we get the image feature library of corresponding dimensions on Corel and Caltech-256 dataset.Then,when doing image retrieval,we use the convolution neural network after migration learning to extract the semantic features of the image to be retrieved,and then reduce the dimension.Finally,the similarity between the characteristics of the retrieved pictures and the corresponding features in the corresponding dimension is compared,and the comparison results are returned,and the returned results are in reverse order.The similarity measure algorithm used in this paper is Euclidean distance.The experimental results show that the convolution neural network after migration learning can effectively separate and express the image features.The convolution neural network after migration learning can learn the advanced semantic features of images.Compared with the traditional content-based image retrieval method,retrieval accuracy is improved at least 3%.After reducing the dimension of the feature,it can reduce the computation of the model and improve the retrieval efficiency,and eliminate the unreasonable or redundant noise in the feature,thus making the image retrieval precision higher than that before reducing the dimension.