Research And Design Of Multi-object Association-based Image Retrieval System

Advances in computer technology and artificial intelligence have led to an explosion in the volume and complexity of multimedia data,and content-based image retrieval systems(CBIR)are widely popular in order to extract useful information from this data.However,the strategy of CBIR systems is to find the "best" representation of visual features and find images similar to the user’s query image based on the selected features,without considering the abstract query.Therefore,this paper designs a system for multi-object association retrieval of vehicle owners and vehicle images using deep learning algorithms.The details of the study in this paper are as follows.1.Implementation of a feature extraction network based on owner and vehicle association images.The Mobile Net network is improved by adding an image pre-processing module,cropping the input image for target detection and adjusting the network parameters.Based on the dataset selected in this paper,the association information of vehicle owner and vehicle images in BIT-Vehicle dataset and Market-1501 dataset are combined and manually annotated.The experimental results show that the model performance is improved by using the YOLO model to acquire image target locations and perform image cropping to remove background interference,and the method of pre-processing the dataset images by data enhancement methods.In addition,the improved feature extraction network proves its superiority in extracting useful information from the images.2.Implementation of multi-object association retrieval based on combined loss function for owner and vehicle images.This paper attempts to mimic the memory mechanism of the human brain,which stores and retrieves images through associative thinking.The human brain can retrieve the complete image from the missing or distorted version of the input image.In addition,by taking a query image as input,the human brain can recall the associated images of that image.For example,if we see an image containing a vehicle,we can associate it with other related images,such as the identity of that owner,the cell phone number of that owner,etc.Firstly,we compare the performance of different triads on non-associative image differentiation and analyze the effect of triad interval parameter on associating positive and negative samples;then we introduce the method of cross-entropy loss and activation function acting together to strongly associate owner and vehicle images;finally,we combine cross-entropy loss and triplet loss to achieve the purpose of strong association and weak association co-remembering,and conduct experimental validation in the dataset.The experimental results show that the semi-hard triplet method can select difficult samples to learn more useful features,and the combined loss function achieves the purpose of associating images.3.Design of an associative image retrieval system.This paper designs and develops a multi-object association retrieval human-computer interaction system based on an improved algorithm with a simple and user-friendly interface.The combination of feature extraction and association algorithm has successfully retrieved the associated images from the database,solving the problem that the traditional image retrieval method has no association storage and retrieval mechanism.