Research On Small Object Detection Algorithm Based On Deep Perception And Hierarchical Attention Mechanism

Small Object Detection(SOD)is a branch of object detection in computer vision,which aims to improve the robustness and accuracy of object detection techniques.SOD faces many challenges,such as a scarcity of samples,low resolution,limited usable features,and difficulty in localization,which need to be addressed to further improve its performance.This article focuses on how to use aggregated features to fully exploit the feature information of small objects and avoid the loss of information caused by downsampling during feature extraction.Based on these challenges,a small object detection algorithm based on deep perception and hierarchical attention mechanisms is proposed.The main research results and work of this algorithm include the following aspects:(1)The MRAM algorithm based on deep perception and hierarchical attention mechanisms.This algorithm uses shallow feature maps for localization information and deep feature maps for high-level semantic information to simultaneously solve the challenges of small object localization and information loss.In this algorithm,an adaptive data augmentation module(ADA),is first proposed to randomly apply image enhancement methods to expand the sample size,enabling the network to have better generalization and robustness.Secondly,a multi-level residual module(MLR),is proposed to extract information of different scales on the same feature map,avoiding the problem of information loss caused by multiple downsampling operations on lowresolution targets.Thirdly,a hierarchical multi-attention mechanism(HMA),is proposed to reassign different weights to the pyramid’s different levels of features in both spatial and channel dimensions,enabling the network to learn aggregated features fully and further improve the model’s robustness.Finally,a cascade detection unit(CDU),is proposed to improve the recall rate of small object detection by using a higher-resolution detection layer and anchor boxes,enhancing the overall performance.Through the combined action of these four modules,MRAM shows excellent robustness in small object datasets.(2)A smart visual analysis platform for security monitoring videos is designed and implemented to apply the MRAM algorithm in practical scenarios and test its effectiveness and feasibility.The smart visual analysis platform for security monitoring videos includes two subsystems,namely,an intelligent analysis and management system for public safety monitoring videos and an intelligent defect detection system for PCB circuit board production.In the intelligent analysis and management system for public safety monitoring videos,the real-time monitoring and analysis of workers wearing safety helmets are implemented using MRAM’s small object detection capability.In the intelligent defect detection system for PCB circuit board production,an innovative combination of MRAM and erosion and dilation operations is used to determine the defect’s location.The defect locations obtained by closing operation and those detected by MRAM are intersected,and the intersection is fed into a classification network to further determine the type and confidence of the defect.In addition,the Head Safe Net safety helmet detection dataset is constructed,which includes 28,417 images and 209,710 data annotations.The dataset images mainly come from images collected by monitoring in industrial scenes.