Research On Image Feature Fusion Method For Pattern Image Retrieval

With the development of textile CAD technology,the number of newly designed printing patterns is growing with each passing day.Facing these massive printing patterns,traditional handcraft retrieval of printing patterns is no longer feasible for textile enterprises.How to quickly and accurately find similar patterns in the design gallery is of great significance to help enterprises greatly reduce the cost of proofing and time,and realize digital production.In recent years,printing patterns have attracted the attention of many researchers due to the diverse styles,wide varieties and fine-grained content.The use of printing patterns is related to intellectual property,so the collection of printing patterns is very difficult.Currently,there are few relevant public data sets.At the same time,compared with the patterns in daily life,the content of printed patterns is more complicated and harder to be retrieved.Under this background,based on the feature fusion method,this thesis conducts in-depth research on the construction of the printing pattern data set and the retrieval of printing patterns.The main work is summarized as follows:(1)A data set of printing patterns is constructed.In view of the various demand scenarios of printing patterns in reality,we summary the characteristics of printing patterns and design the construction process and method of printing pattern data sets.Ultimately,a printing pattern data set which is called Pattern data set is constructed for multiple task scenarios.The Pattern data set contains multiple subsets.Firstly,according to the nature of the pictures,it can be divided into design pictures,scan pictures and photos.Secondly,according to whether the pictures are cropped,they can be divided into original pictures and cropped pictures.Thirdly,according to different tasks,they are divided into pattern retrieval sets,pattern detection sets,etc.We verifies the usability of the data set for pattern retrieval,pattern detection and other tasks through experiments.(2)A pattern retrieval model called pattern global and local feature network(PGLN)based on feature fusion is proposed.Aiming at the problem that a single feature cannot fully express the content of a picture,this thesis proposes a feature fusion method and constructs a deep network model called PGLN that combines global and local features.Among them,the global feature uses the local pooling feature map of the deep network to efficiently integrate the salient features of the input image,making the network more robust to changes in the input image;local feature branches use the attention mechanism and the interactive feature layer to detect the salient area of the image.In order to verify the effectiveness of the PGLN model,multiple sets of comparative experiments are carried out on the Pattern data set.Experiments show that the fusion feature can achieve better retrieval results compared with a single feature.At the same time,compared with other feature fusion algorithms,the PGLN model has achieved the best retrieval effect on the retrieval task of the Pattern data set.