Study Of Improved Feature Pyramid Networks Based Algorithms For Detecting Small Objects

In real life,object detection has many applications with significant practical value.With the rapid development of deep learning,the accuracy of object detection has repeatedly hit a new high.However,due to some characteristics of small-size object,improvement of detection for small targets has always been a diffculty,which has gradually become one of the most challenging task.Aiming at some issues in small object detection,the following works have been done in this paper.An image enhancement and augmentation method based on small object data set is adopted.In some classical object detection algorithms,the small objects often lead to the problem of unbalance between positive and negative samples in detection.The data preprocessing method adopted in this paper can overcome this problem to a great degree.Appropriate small sample pasting can greatly improve the accuracy of small object detection without changing the network structure.Feasibility and effectiveness of the method have been proved by multiple combination experiments.Combing the concatenation module with channel attention module to modify the network connection is proposed based on the feature pyramid network.In this way,the high level semantic features and low level detail features can be integrated more efficiently.The concatenation module has bigger search space which makes training easy.On the one hand,the improved channel attention module is more sensitive to small targets,on the other hand,it is beneficial to feature weighting.We not only demonstrate the superiority of the new network in theory,but also prove the superiority of the new network in the subsequent experiments.Based on the concatenation feature pyramid network mentioned above,this paper innovatively proposes the feature matrix transformation module and high level feature augmentation module,which replaces the up-sampling step and 1×1 convolution operation in the feature pyramid network before.It avoids introducing of noise during the information amplification and the information loss during dimension reduction.The high level feature augmentation module makes up the deficiency that the high level feature is not fully utilized,which further improve the performance of small object detection.