Multi-scene Small Object Detection Via YOLOv4

Applications of small object detection have a wide range of scenarios and this topic is also the research difficulty of object detection and recognition.Therefore,improving the accuracy of small object detection is not only important in theory,but also significant in practice.However,the current object detection algorithms are not effective in the small object detection tasks.In order to improve the detection accuracy of the task,a type of generalized improved algorithm based on YOLOv4 model is proposed.We combine spatial attention with channel attention networks to enhance the weight of object feature map and name it Mixed Attention Network(MA).The improved algorithm is called YOLOv4 with Mixed Attention(YOLOv4-MA).In order to solve the problems of missed and false detection in small objects,this paper modifies the network structure of YOLOv4 to improve the detection accuracy of small objects.YOLOv4-MA has been applied in Person Re-identification and Face Recognition fields.This paper mainly includes the following contents: feature extraction and feature fusion.Firstly,this paper introduces the research methods and significance of small object detection,as well as the research status at home and abroad.From the perspective of feature extraction,in order to solve the difficulties of weak expression ability,sample imbalance and uneven distribution in small objects,we propose a novel attention model MA to improve the weight of small objects in the feature graphs.At the same time,we improve the robustness and detection accuracy of the model by optimizing Mosaic Augmentation Algorithm and secondary clustering of small object data sets with Kmeans + + Algorithm before training.Finally,we use Focal and Efficient Intersection Over Union(Focal-EIOU)Loss Function to replace Complete Intersection Over Union(CIOU)to alleviate the gradient explosion problems in the process of small object features transfer.We carry out a large number of comparative experiments on Pascal VOC and Visdrone data sets to verify the effectiveness of YOLOv4-MAAlgorithm in this paper.Secondly,in addition to the above contents,we optimize the YOLOv4-MA Network from the perspective of feature fusion.Because small objects do not exist alone,they are often associated with the background and other objects.Therefore,we design a kind of Dilated Convolution Module(DCM)with larger receptive field to mine the association information of feature maps.At the same time,Double Deconvolution Module(DDM)is added to the neck of YOLOv4-MA Network to improve the fusion ability of semantic information.Finally,we use Attention-guided Context Feature Pyramid Network(AC-FPN)with better performance to replace the FPN in Path Aggregation Network(PANet).On the augmented Pascal VOC and Visdrone data sets,we carry out a large number of comparative experiments to prove the effectiveness of the optimization scheme.