Research On Small Object Detection Algorithm Based On Second-order Dynamic Convolution Network

Small object detection is one of the difficult problems in current object detection field.Since the low resolution of small objects(objects size smaller than 32×32 pixels),which contain less pixel information,it is easy to be affected by irrelevant information such as background and noise.It is difficult to model effective and discriminative small object features when convolutional sampling is carried out by the current generic detection algorithms,which leads to unsatisfactory detection precision and localization of small objects.In view of the efficient feature learning ability of dynamic convolutional networks and the advantage of second-order statistical modeling methods to capture the higher-order representation information of features,the thesis proposes a small object detection algorithm based on second-order dynamic convolutional networks.The main research contents of the thesis are as follows:1.The feature extraction network,dynamic convolutional network and second-order statistical modeling methods are studied and analyzed in terms of their development & working mechanism and application advantages.Then take the typical object detection models as the target of statement and summarize the architecture characteristics of different model frameworks from point to surface,and the solutions adopted for the small object detection problem.2.A feature optimization strategy combining dynamic convolution networks and secondorder statistical modeling methods is proposed,and So DNet(Second-order Dynamic Network)feature extraction network is constructed.Based on the Res Net50 and CSPDarknet53 networks,the standard convolution operation in the residual unit is replaced by the Dynamic Convolution(Dy-Conv)operator,and the global second-order pooling(GSo P)module is used to estimate the covariance of the obtained dynamic features and generate object features with higher-order(second-order)information representation,so as to achieve the purpose of strengthening the network representation ability.The experimental results show that compared with the original network,the classification accuracy of So D-Res Net50 and So D-CSPDarknet53 feature extraction networks on the VOC2007 test dataset is improved by 2.71% and 3.02%,respectively.3.Based on So D-CSPDarknet53,a small target detector with position attention and adaptive receptive field characteristics is designed.Taking YOLOv4 as the baseline model,in view of the weak location information of small objects and the unreasonable matching between receptive field and object scale,the position attention module(PAM)and adaptive receptive field(ARF)module are constructed to optimize the detection network.PAM uses three parallel1×1 convolution kernels and Reshape & Transpose operations to process spatial information encoding and dimension transform of feature maps.Then,the potential feature spaces mapped by two branches are fused,and the position attention feature maps are output by continuous average pooling and non-linear activation layer.It then interacts with the original feature space on another branch.ARF is mainly composed of 3×3 dilated convolution,which for adjusting the size of receptive field.Finally,this module is embedded into each detection head for multiscale prediction.On the Vis Drone aerial datasets,experimental results show that compared with the baseline model YOLOv4,the So DNet-based small object detection algorithm has improved the overall mean average precision(m AP)by 5.54%,of which the m AP of small objects has increased by 3.55%.