Research On Lightweight Object Detection Algorithm Based On Multi-scale Feature Fusion

Object detection is one of the most important research directions in the field of computer vision,which is widely used in face recognition,pedestrian detection,remote sensing image detection and other fields.The current object detection algorithms achieve excellent detection performance for large-scale and medium-scale objects,but it is difficult to accurately detect small objects in real complex scenes.There are problems with false detection and missing detection.This is because the small objects have few pixels and low resolution.At the same time,small object detection is easy to be interfered by background factors such as illumination,blur and occlusion.Therefore,this thesis focuses on small object detection and improves the detection performance of small objects from the aspects of feature enhancement,feature fusion and anchor optimization.Specific research work is as follows:(1)A lightweight multi-scale network-based object detection algorithm is proposed to solve the problem that features are easy to be lost after down-sampling small objects and the feature expression ability is weak.This algorithm makes full use of multi-scale information to improve the detection effect of small objects.Firstly,based on the basic framework of Mobile Netv2-SSD,a multi-scale feature fusion module is proposed to fully integrate high-level and low-level features and effectively improve the feature expression ability.Secondly,a lightweight receptive field enhancement module is designed to effectively increase the receptive field of the feature maps by dilated convolution with different dilated rates.Finally,in the process of multi-scale prediction,an efficient channel attention module is introduced to improve the information correlation between feature maps and make full use of object context information.The experiment results show that the m AP of LMSN on PASCAL VOC and RSOD datasets reaches 75.76% and89.32%,respectively,which is 5.79% and 11.14% higher than Mobile Netv2-SSD.At the same time,it maintained detection speeds of 61 FPS and 64 FPS.This proves the effectiveness of LMSN for small object detection.(2)A balanced reverse feature fusion network-based object detection algorithm is proposed to solve the problem of insufficient utilization of detail and semantic information in feature maps.Based on the basic structure of Mobile Netv2-YOLOv4,three improvements are proposed for the characteristics of small objects.Firstly,an efficient convolutional block attention module is designed to integrate small object features in both channel and space dimensions,which strengthens the network’s focus on small object features and suppresses the irrelevant background interference.Secondly,a reverse feature pyramid network is proposed to fusion the features of adjacent and nonadjacent levels to improve the utilization rate detail and semantic information in feature maps.Finally,the K-means++ algorithm is applied to cluster analysis to generate the anchor suitable for small object detection,so as to improve the positioning accuracy of small objects.Experimental results show that the m AP of B-RFFN on PASCAL VOC dataset reaches 81.0% and the detection speed is 61 FPS.The m AP on the RSOD dataset is 92.66% and the detection speed is also maintained at 65 FPS.It is proved that B-RFFN can effectively improve the accuracy of small objects while maintaining competitive detection speed,and better balance the speed and accuracy of object detection.