Research On Object Detection Method Based On Multi- Scale Refine Fusion Network

Object detection aims to recognize and localize each object instance in an image.Object detection is the core of computer vision,which is widely used in image recognition,large-scale scene recognition and so on.Deep learning methods are widely used in object detection,and have achieved far better results than traditional methods.The region-free methods and the region-based methods are mainly two kinds of object detection methods based on deep learning.The region-based methods generate many candidate areas on the picture and use the area that may contain objects as the recommended area.Region-based methods usually have higher accuracy and robustness.Therefore,this paper mainly studies detection methods based on region-based methods.The region proposal network(RPN)is a commonly used region-based methods.It extracts features by convolutional neural networks,and then uses the feature information to generate high-quality suggested regions.RPN improves the accuracy and detection speed by reducing the number of candidate areas.However,since the single scale feature maps contains insufficient information,the ability of RPN to detect and locate small-sized objects is poor.On the basis of summarizing the Multi-scale method and attention mechanism,this paper proposes a region proposal network that can efficiently use multi-scale information.The proposed method is called multi-scale region proposal network(MS-RPN).Considering above problems and situations,the main work and innovation of this dissertation are as follows:(1)This paper studies the multi-scale method and proposes a feature fusion network based on the multi-scale method.Feature fusion network integrates multiple scale features with different sizes and different distributions,which improves the problem of single scale feature maps contains insufficient information.(2)Based on the feature fusion network,a feature refinement network based on the attention mechanism is added.Feature fusion network can obtain the probability distribution of attention by calculating the correlation between features from different scale.The purpose is to give different weights to different features,and then reduce the dimension of weighted features.In this way,the information redundancy caused by feature fusion is solved and important features are highlighted,which improves the quality of features and improves the effect of objectdetection.(3)In order to verify the effectiveness of this method,a comparative experiment is performed on two authoritative target detection datasets,and the method is applied to the actual application data set to verify the practicality of it.Experimental results show that MS-RPN can achieve better performance on all data sets.Combined with practical applications,this paper improves non-maximum suppression,and proposes adaptive nonmaximum suppression(ANMS),which further improves the practicality of the detector.