An Improved SSD Object Detection Algorithm Based On Multi-layer Feature Fusion

As an important research direction in computer vision,object detection technology is widely used in intelligent traffic,image retrieval,intelligent monitoring and other fields.Most of the traditional object detection algorithms are based on artificially designed features and small sample training,so they are easily affected by many factors such as motion blur,morphological changes,environmental changes,and object occlusion.Therefore,the detection performance of traditional object detection algorithms is unsatisfactory.With the deep convolution activation feature in the image classification field has made breakthrough,the entire academic and industrial circles have extensively used deep learning to achieve object detection.Although the object detection algorithm based on deep learning has greatly surpassed the traditional algorithms,there is still a big gap away from the level that people can achieve,and the existing object detection algorithms still have much room for improvement in speed and accuracy.For this reason,it is a very valuable work to study an object detection algorithm with high detection speed and high detection accuracy.Therefore,based on previous research work,this thesis puts forward some improvement strategies and sums it up as follows:1.The advantages and disadvantages of SSD(Single Shot Multibox Detector)algorithm are analyzed in depth.Considering that the feature map used for prediction in the SSD network is not reused,the AFFSSD(adjacent feature map fused SSD)is proposed.AFFSSD uses the SSD network as the basic framework,and deconvolutes some high-level feature maps,and then fuse the low-level feature maps with "element-sum" operations to form new feature maps.According to the characteristics of the selected fusion feature map,AFFSSD is divided into AFFSSD(7&4)and AFFSSD(5&4),and their mean average precision on the PASCAL VOC2007 dataset is 78.9% and 79.1%,respectively.On the MSCOCO dataset,the AFFSSD(5&4)has a mean average precision of 8.9% for small targets,which is higher than SSD and DSSD(Deconvolutional Single Shot Detector).2.In order to further enhance the semantic information of feature maps,the TFFSSD(twice feature fused SSD)is proposed.The TFFSSD first performs downsampling of the lowlevel feature maps that have been fused through the feature pyramid network.Then they are fused with high-level feature maps,so that high-level feature maps also contain the semantic information of low-level feature maps.Experimental results on the MSCOCO dataset show that the mean average precision of TFFSSD in medium and small objects is increased by 0.3% and 0.6% respectively compared with AFFSSD(5&4).