Research On Attention Mechanism And Multi-scale Feature Fusion Method For Object Detectio

Object detection is a fundamental task in the field of computer vision,and it plays a crucial role in areas such as autonomous driving,video surveillance,and face recognition.In recent years,the advancement of deep learning technology has promoted the development of object detection,however,there are still some unconquered challenges in the research of object detection algorithms,such as the problem of detector performance degradation brought by the scale diversity of instances in images and the problem of poor detection effect in complex object detection scenarios.To alleviate these problems,this paper first investigates the multiscale feature fusion method based on the two-stage object detection algorithm,and then explores how to combine the multiscale feature fusion method with the attention mechanism and consider the speed-accuracy tradeoff on the anchor-free object detection algorithm based on this.The main contents and contributions of this paper are as follows:(1)Multi-scale feature fusion pyramid network for object detection: To address the impact of scale variation of instances in images on detector performance,this paper proposes a generalized multi-scale feature fusion pyramid network.The network has a bottom-up feature enhancement mechanism to produce a feature representation that is robust to scale variations.The method proposed in this paper borrows the feature pyramid network structure,but with deeper and wider convolutional layers,and is plug-and-play.In addition,to further enhance the information flow of the network,a skip enhancement network module is designed to aggregate the texture features in the shallow layer of the network as well as the semantic features in the deep layer of the network.The experimental results show that the proposed method not only achieves further improvement in several object detection algorithm benchmarks,but also can be applied to tasks such as instance segmentation.(2)Anchor-free object detection algorithm based on multi-scale feature aggregation attention: In complex scenes,where the scale of instances in the image is small and contains a lot of background noise,more comprehensive image features need to be extracted in order to further improve the performance of the detection algorithm.To this end,this paper proposes an anchorless object detection algorithm based on multiscale feature aggregation attention,which tries to combine multiscale feature fusion methods with visual attention mechanisms and introduce them into anchorless object detection algorithms,taking into account the detection speed while improving detection accuracy.In this paper,a multi-scale feature aggregation attention network is designed to achieve the purpose of aggregating the features of different perceptual fields of the backbone network while focusing on the important regions of the feature map.In addition,this paper also designs a key point prediction network decoupled from the regression branch,so as to further improve the localization accuracy of the algorithm.The experimental results show that the algorithm can effectively suppress the influence of background noise and small target instances on the detector and improve the accuracy of the objective detection algorithm in complex scenes under the premise of ensuring real-time detection.