Research On Object Recognition Optimization Of Hotel User-generated Images Based On Deep Learning

With the rapid development of e-commerce and Internet technology,more and more people choose to consume through the Internet and are willing to share their comments on the experience of using goods or services on the online platform.Through the mining and analysis of online reviews,it can provide other consumers with multi-dimensional product experience information and help potential consumers to make better purchasing decisions.At the same time,it also enables merchants to further understand the advantages and disadvantages of their products,so as to optimize their management strategies,improve the quality of products or services,and strengthen their competitive advantages.At present,researches related to online reviews have attracted extensive attention from scholars at home and abroad,and the researches on online reviews in the field of hotels and tourism have been gradually deepened.However,the existing researches mainly focus on the analysis of the text content,emotional tendency of the text and the rating of the reviewers.User-generated pictures are an important part of online comments.Objects contained in pictures can fully demonstrate the characteristics of products and services,and at the same time reflect the degree of concern of users on product attributes.Objects contained in pictures are of great value for further research on online comments.However,at present,it is rare to find the object information contained in picture comments,which is mainly based on content analysis and semiotic analysis.These methods cannot provide effective feature extraction methods for pictures in hotel scenes.Therefore,based on the deep-learning target detection algorithm Faster R-CNN and combining the characteristics of the hotel user generated images application scene,this paper analyzes the reasons why the identification accuracy of Faster R-CNN is low in this scene,and proposes targeted optimization measures to meet the detection requirements of the hotel user generated images application scene.Combining the characteristics of the target objects in the user-generated images of hotels,such as diverse types,varying shapes and sizes,and unbalanced sample quantity,this paper proposes a more efficient target detection model,Faster R-CNN-FFS,suitable for the hotel scene.Specific research work includes:First,this paper takes the image data obtained from user-generated comments on e-commerce platforms such as "Qunar" and"Ctrip" as the data support,and makes a total of 14,901 images of eight types of objects on the hotel user-generated image data set.Then,aiming at how to effectively alleviate the problems such as false detection and missing detection of multi-object detection and small-size target identification in complex and unconstrained hotel scenes,this paper improved the feature extraction network structure of the Faster R-CNN model and proposed to introduce the feature fusion structure into the Faster R-CNN for feature extraction.Then,in view of the imbalance between positive samples,negative samples,difficult samples and simple samples in the training process,Focal Loss function was proposed to replace the cross entropy Loss function in the original model.Secondly,the candidate box screening mechanism based on NMS algorithm in the existing model is relatively simple and direct,which leads to a low score of missing selection but can correctly represent the candidate areas of another object.In this paper,soft-NMS is used to optimize the non-maximum suppression algorithm of Faster R-CNN to select potential target regions with better quality for regression.Finally,the Faster R-CNN-FFS model proposed in this paper was verified,and the recognition effect of the model proposed in this paper was significantly improved,with the final mAP value reaching 69%.The above work proves that the research work of this paper fully considers the data characteristics of object detection of user-generated images in the field of hotels and effectively solves the problem of multi-target identification in complex unconstrained scenarios.The results for the study of the target object detection under the complicated scene provides a new train of thought,as well as development based on the modal of structured data(score)and unstructured data(text and images)analysis of online reviews research provides effective theoretical support,further enrich the characteristics of user generated content,for further research on hotel online reviews provide effective support,to optimize service and hotel management decisions have important reference value.