Research On Small Object Recognition Algorithm Based On SE-SSD Network

In recent years,with the continuous research of deep learning algorithm,especially in thefields of face recognition and intelligent driving,deep learning algorithm has gradually become practical.Deep learning algorithms and their applications have gradually become an indispensable part of our lives.How to accurately recognize the target in the given environment where the size of multiple objects vary greatly,on the condition that the network ensure a certain detection speed,which is the main task in the design of neural network architecture in recent years.Therefore,the small object recognition algorithm based on SSD neural network has also become one of the current research hotspots.This paper aims at the defects of poor accuracy of traditional SSD networks when detecting and recognizing small-sized objects,and in order to further improve the recognition performance of the network,we first proposed an SSD network(SE-SSD)based on the "Squeeze-and-Excitation" module.As a currently popular method,increasing the number of single-layer feature maps or the depth of neural network architecture can effectively improve the performance of deep networks.This paper is to introduce SE modules and analyze how to effectively use and optimize SE modules to improve SSD performance.Compared with the traditional method of increasing the depth of the CNN architecture,the network proposed in this paper explicitly models the nonlinear relationship between feature channels,and selectively emphasizes or suppresses the feature information of different channels to effectively improve the network performance.Based on the SE-SSD network architecture,this paper further proposes two small object recognition optimization methods: SE-SSD optimization network based on feature pyramid network(FPN)and small object recognition strategy based on data augmentation.The former introduces the architecture of FPN feature extraction network to perform multi-layer feature fusion on the feature maps involved in classification prediction in SE-SSD network.This method not only enriches the context of small object features,but also realizes the two-way transmission of feature information upstream and downstream,thereby improving the ability of SE-SSD network to identify small objects.The latter starts from the perspective of the dataset,analyzes and proves that there is a significant gap between the total number and the number of single object matching anchors between the calibrated object of different scales ofthe mainstream MS COCO dataset and Pascal VOC dataset.Therefore,this paper uses two methods of image oversampling and object paste-copy to achieve data augmentation.We analyzes and discusses how to effectively paste small objects.This method allows us to weigh the detection quality of large and small objects.The experimental results of this paper are all tested and evaluated on mainstream datasets:the feature fusion network based on SE-SSD has 78.9 m AP and 28.5 m AP on the VOC2007 and MS COCO2014 datasets respectively,and compared with the SSD network,the accuracy of small object recognition is improved by 37%;SE-SSD based on the data augmentation obtains 78.8 m AP on the VOC2007 dataset and improves the accuracy of small object recognition from 33.7 m AP to 35.6 m AP while guaranteeing the model operation speed 57.3 fps.