Development And Application Of Product Recognition Algorithm Based On Image Retrieval

Product identification is one of the core technologies in the modern retail industry.Compared with traditional product identification methods,product recognition based on computer vision has the advantages of fast,accurate and low cost,which has broad application prospects.Most modern algorithms for product image recognition mainly use image classification networks.In practical applications,there are problems such as being unable to change product categories flexibly,limiting the number of product categories,and lacking of category generalization ability.In order to recognition products with high speed and accuracy,while avoiding the above problems,this thesis developes a product recognition algorithm based on image retrieval that has research significance and engineering application value.The product recognition algorithm in this thesis contains two phases:feature extraction and feature matching.Firstly,this thesis designs and implements the feature extraction network based on metric learning.Using pre-trained backbone extract rich global information from product image,and then mapping the information into output feature vector.Improve the triplet loss and use it to supervise feature extraction,which makes the networks generate more discriminative features.In addition,Traning with classification loss helps to train the network faster and more stable.Then develops the feature extraction network based on multi-branch structure.The designed multi-branch structure splits and maps the global feature extracted by backbone to encode different image information into multiple feature vectors.Combining these features vectors for feature matching improves the robustness and accuracy of the model.Finally this thesis designs the feature matching algorithm,which combines the compute and ranking of the cosine similarity between features to match the query products with the registered products and then obtain the recognition results.Testing on 36995 images of 100 products,the accuacy of the model based on metric learning is 98.65%and recognition only consumes 15.13ms for an input image with 224×224 resolution,then the model based on multi-branch structure can achieve 99.14%accuracy although recognizing one image takes 31.6ms.This thesis applies the above product recognition algorithm to the shelf scene and develops a shelf product detection and recognition algorithm.According to the characteristics of shelf products,such as being densely placed and existence of many small objects,this thesis made four improvements to Faster R-CNN to adapt to shelf product detection firstly,including changing the backbone,using the feature pyramid architecture,replacing RoIAlign with RoIPool,and optimizing the anchor setting based on K-means.Then improve the feature extraction network based on the multi-branch structure by adding the feature compression branch to output low-dimensional feaure vector.After that perform feature matching with the low-dimensional feature vectors and high-dimensional feature vectors generated by multi-branch structures,which can push shelf product recognition to higher speed and accuracy.After analysis,the precision of the improved Faster R-CNN is 99.21%and the missed detection rate is only 1.37%.The improved product recognition algorithm can get a accuracy of 99.29%in the 448 products and recognition only consumes 17.1 ms for per image.