Object Detection Based On Information Fusion And Sample Dynamic Assignment

Object Detection is one of the fundamental and significant research directions in computer vision.Given an image,the task of object detection algorithms is to classify and localize the objects contained in the image.Although researchers have achieved fruitful accomplishment in the field of object detection,there are still a great number of challenges in object detection to be overcome.First,there are various interference factors in real-life scenarios such as varied size of different objects,the length-width ratio of different objects,occlusion,blur,which puts forward higher requirements for the robustness and precision of object detection algorithms.Second,accurate and efficient object location requires the object detection algorithm to have stronger feature discrimination of objects and border location ability.This paper concentrates on the research of object detection and main contribution can be summarized as follows:1.An object detection algorithm based on layer-level semantic information aggregation of feature pyramid network is proposed.In order to fuse the semantic information of different layers of feature pyramid network,a hierarchical semantic information aggregation network is constructed.The network aggregates the rich semantic information of different layers and captures the relationship among objects and between the objects and background.Thus,the accuracy of object detection algorithm is improved.Aiming at the problem of information loss during feature fusion,a dense excitation feature enhancement module is constructed.The module enhances the feature representation of the objects and makes the network focus on the differences between objects and suppress background areas.In other words,the module strengthens the saliency of object areas and weakens the impact of background areas on detection performance,which can improve the accuracy of object detection.2.An object detection algorithm based on multi-layer region of interest feature fusion is proposed.Traditionally,heuristic proposals assignment strategy will assign proposals of similar size to different layers.We construct a novel proposals assignment strategy,in which each proposal is assigned to all feature layers so that each proposal can obtain adaptive feature.Aiming at the problem that the high-level convolution features lack detailed information and low-level convolution features lack semantic information,a multi-layer region of interest feature fusion network is constructed.The network can fuse the semantic information and detail information of multi-layer region of interest features.Since convolutional neural network mainly focus on spatial dimension information and ignores channel dimension information,a channel and spatial attention fusion mechanism is constructed.The mechanism enables the network to pay attention to both spatial and channel information,which improves the performance of object detection algorithm.3.An object detection algorithm based on dynamic assignment of positive and negative samples and edge-aware network is proposed.Aiming at the problem that sampling strategy of positive and negative samples with fixed threshold does not match the dynamic network training,a dynamic assignment strategy of positive and negative samples is proposed.The proposed strategy can make sampling of positive and negative samples changes dynamically and the quality of proposals become higher with the training of the network,which improves the training effect of the classification network and the performance of object detection algorithm.It is difficult to regress when the coordinates of low-quality candidate boxes are far from real coordinates.By integrating the proposed dynamic assignment strategy of positive and negative samples into the edge-aware network,we improve the quality of the proposals in regression network,reduce the difficulty of proposals regression,and improve the classification performance.Finally,the accuracy of object detection is improved.