Research And Implementation Of Dense Small Target Detection Optimization Algorithm Based On SSD

At present,the scale of the target and the density between the targets are the key factors that affect the performance of target detection technology and the main challenges faced by the research of target detection algorithms.Focusing on the detection of dense small target,an optimized algorithm based on the SSD algorithm that has balances detection accuracy and efficiency is proposed in thesis.The optimization design covers several aspects such as data augmentation,feature enhancement,loss function,and post-processing methods.The main research content and achievements in thesis include the following aspects:(1)Optimization design of data augmentation method based on Cut Mix.The number and proportion of small target samples in commonly used datasets are usually low.Therefore,using the Cut Mix method that combines information and labels to augment data in thesis.To avoid the common problem of single-image fusion and random size scaling,proposing optimization method in multiple target fusion and fixed image size in thesis.Experiments show that m AP is improved by 0.8% on PASCAL VOC dataset,and AP50:95 is improved by2.2% on MS COCO dataset.(2)Optimization design of feature enhancement method and feature fusion method.Small targets often lack key detailed feature information due to their small scale.Therefore,based on data augmentation,optimizing the network architecture of the SSD algorithm by adding dual mechanism feature enhancement and bidirectional feature fusion modules in thesis.The dual mechanism feature enhancement module introduces a parallel attention mechanism for weight allocation,which enables the model to better focus on the target area instead of the background.At the same time,the multi-core grouping convolution mechanism is used to seek the optimal convolution ratio,expand the receptive field of the feature layer,achieving a balance between detection speed and semantic information enhancement.The bidirectional feature fusion module constructs bidirectional network channels,and strengthens the connection between feature layer detailed appearance information and high-dimensional semantic information through element-wise addition.Experiments show that the algorithm optimized by(1)and(2)is integrated achieved a 4.7% increase in m AP in the PASCAL VOC dataset and a 7.5% increase in AP50:95 in the MS COCO dataset.(3)Optimization design of loss function and post-processing method.To solve the problem of missing detection caused by mutual suppression of dense targets due to the close distance of detection boxes,replacing the original regression loss of the algorithm with an loss function based on Repulsion optimization in thesis,mainly adjusting the parameter setting that measures the relative distance in the attraction term,enabling the threshold of the repulsion term,and selecting repulsion objects.This optimization aims to make the detection boxes generated by the algorithm as dispersed as possible for dense targets.At the same time,the paper further optimizes the original post-processing method by building a mapping relationship between the threshold and the number of detection boxes,achieving adaptive threshold selection.Experiments show that the optimization algorithm after integrating the three optimization designs proposed in thesis achieved a 5.4% increase in m AP in the PASCAL VOC dataset and an 8.8% increase in AP50:95 in the MS COCO dataset compared with the original algorithm.