Research On Image Retrieval Method Of Museum Relics Based On Deep Residual Network

With the rapid development and application of computer technology in the field of multimedia,there are more and more cultural relics images,which have become important data resources for the construction of smart museums.However,in the current investigation of the museum,it is found that these image data have not been used efficiently,especially in image retrieval,they still remain in the original stage of manual labeling,which is carried out by manual labeling.When the image quality is poor or the image scale is large,the effect of manual annotation often appears to be insufficient,relying too much on manual,lacking certain flexibility,and the quality of information annotation is also uneven.Therefore,aiming at the above problems,this thesis proposes a deep learning based image retrieval method for museum relics.The main work of this thesis is as follows:(1)In terms of image data,we built our own cultural relic data set through Internet crawler and other means,including bronze ware,pottery and porcelain in 3 categories and11 sub-categories,400 kinds of cultural relics in each category,about 4 images in each cultural relic,4780 images in total,and obtained the original data set.(2)In image preprocessing,an improved Deep Lab V3 + network background segmentation model of cultural relic images is proposed.The depth detachable convolution operation was used to replace the ordinary convolution operation in the process of void space convolution pooling(ASPP)in the original model.Meanwhile,the three-layer depth detachable convolution was used to separate channels in the decoding stage,instead of the one-layer common convolution used in the traditional encoding stage,so as to enhance the depth information acquisition of images.The experimental results show that the improved method is smooth and clear,and the average segmentation time is low.When applied to the current self-built cultural heritage image database,the average time consumption of this method is reduced by 4.7s compared with Grab Cut model,and 0.76 s compared with Deep Lab V3 + model.(3)In image retrieval,an improved deep residual network model based on Res Net50 is proposed.In the structure of residual network,the global pooling method is adopted and the ib N-NET idea is added to construct residual block of benchmark network RESNET50-IBN-A to improve the generalization ability of the model.In addition,the first maximum pooling layer in Res Net50 structure was removed,and then an average pooling layer was added as a transition layer before the specific residual module convolution kernel was used for down-sampling to further reduce the complexity of residual network model and integrate the spatial information of images.Compared with Res Net50,the overall retrieval accuracy is about 2.6 percentage points higher.(4)A relic image retrieval system based on residual network is designed and developed,which includes four modules: relic overview,user management,image management,image retrieval and classification.Aiming at the problems encountered in image processing of the museum’s existing system,the system adopts an improved algorithm to enhance the accuracy of image retrieval and achieve better data administration.