Research On Object Detection Algorithm Based On Graph Neural Network

As an important branch in the field of computer vision,object detection extracts extensive valuable semantic information from images.It has been widely investigated in intelligent monitoring,medical imaging,intelligent transportation and other fields,and has attracted the attention of growing researchers and scholars.Object detection task not only determines the localization of each object in the input image,but also realizes the classification of each object.In recent years,with the rapid development of artificial intelligence,object detection algorithms based on deep learning have emerged.Although many research works on object detection in recent years have improved the performance of detection algorithms by utilizing relational representation in images,these methods fail to fully explore the correlation between objects and labels in images,which limits the improvement of the algorithm.Aiming to address this issue,a novel object detection algorithm based on graph relationship reasoning is proposed in this paper.Based on this method,the design of the relationship representation in the image is improved.An object detection algorithm based on the attention mechanism is proposed.The research content and results of this paper are as follows:(1)Object detection algorithm based on graph relational reasoning is proposed.Existing algorithms usually only focus on correlations within the region proposals or within the labels and pay less attention to the relationships between them.To solve this problem,this paper proposes a novel object detection framework based on graph relational reasoning.At first,the framework constructs region proposals and label embeddings into a uniform relational graph.Then,an improved graph convolutional network is developed to perform relational reasoning on the relational graph.Finally,the labels of region proposals are predicted from the relational graph,and the features of the updated region proposals are extracted for bounding box regression.The proposed algorithm further enhances the performance of object detection.(2)Object detection algorithm based on attention mechanism is proposed.Lots of existing object detection algorithms based on graph convolution network widely rely on hand-crafted graph structures,which may introduce unreliable relationship on the graph and seriously affect the performance of object detection.To solve this problem,this paper proposes an object detection framework based on attention mechanism.Firstly,the extracted region proposals and label word embeddings are taken as two independent sets.Then,a self-attention module is designed to discover the inner-relationships within the two sets.In addition,a cross-attention module is developed to learn the inter-relationships between the two sets.Finally,the classification results are predicted according to the similarity of the two updated sets,and the updated features of region proposals are extracted for bounding box regression.The performance of object detection is improved.In this paper,the proposed frameworks are embedded into various mainstream detection algorithms.Extensive experimental results on two public datasets have demonstrated the effectiveness and robustness of the proposed methods.