Research On Small Object Detection Algorithm Based On SSD

With the development of deep learning technology and the improvement of GPU computing power,object detection has been widely used in various fields such as agriculture and industry.There will be a large number of small objects in actual detection scenarios such as unmanned driving and industrial defect detection.However,the small object has a small size,occupies few pixels,and is not easy to detect,which has always been a difficulty in the field of object detection.In this thesis,a deep learning-based SSD(Single Shot Multibox Detector)model is selected as the basic model.The purpose of this study is to improve the small object detection effect of the SSD model.The SSD model is improved from three aspects: network structure,sample balance,and receptive field.The main work and innovations of this thesis are as follows:(1)Aiming at the problem that the SSD model has insufficient ability to extract small object features,from the perspective of network structure,an improved SSD algorithm based on cascading FPN and ASE modules is proposed.First,use the cascading FPN module to perform feature fusion on different layer feature maps to generate two feature pyramids,then use the ASE module to merge the two pyramids and obtain the importance between different channels of the feature map by learning,different channels are given different weights.Through experimental verification,the cascading FPN and ASE modules are added to the SSD model to solve the problem of insufficient capacity of the SSD model to extract small object features.(2)Aiming at the problem of sample imbalance in the data itself and the SSD model training process,based on the improved SSD algorithm based on the cascaded FPN and ASE modules,three sample balance methods are proposed: small object data augmentation,phased IOU threshold,and improved Focal Loss.Use the small object data augmentation method to solve the problem that the number of small objects in the data set and the number of pictures containing small objects are smaller than those of large and medium objects.Use the phased IOU threshold method to set different IOU thresholds for different sizes of priori boxes,appropriately reduce the screening conditions of small object prior boxes,increase the number of priori boxes that pass through the screening,and solve the problem of sample imbalance during training.The improved Focal Loss adds dynamic weights to Focal Loss according to the size of the priori box,further increasing the bias of small objects during model training,and solving the problem of sample balance from the side.Through experimental verification,the three sample balance methods have improved the detection effect of small objects to varying degrees,and there is no mutual exclusion among the three methods.(3)Aiming at the problem of low feature recognition of the small object itself,from the perspective of the receptive field,the RFB module is inserted into the improved SSD algorithm based on the cascaded FPN and ASE modules.The multi-scale dilated convolution in the RFB module is used to fuse features of different receptive fields to enrich the context information of small objects.According to the different insertion positions and numbers of RFB modules,multiple sets of comparative experiments are designed.Through experimental verification,the model can enhance the recognition of small object features and improve the detection effect of small objects after being inserted into RFB module.The insertion positions and numbers are different,and the degree of improvement of the detection effect is also different.