Research On Small Object Detection Method Of Safety Helmet Based On Deep Learning

Wearing safety helmets is one of the most effective protection measures for workers against accidental head injuries while performing construction work.At present,the use of object detection technology to replace traditional supervision methods to achieve automatic monitoring of safety helmets has become a hot research problem.Considering the important application significance of safety helmet object detection in the field of industrial production safety monitoring,this paper conducts research on the problem of small object detection of safety helmets in construction operation scenarios based on the deep learning-based object detection method,as follows:1.Safety helmet detection method based on context extraction and attention feature fusionThe existing object detection methods are insufficient for detecting small-scale safety helmet objects in surveillance image scenes and cannot balance the detection effect and speed.We propose a safety helmet detection method based on context extraction and attention feature fusion.The method enhances the semantic learning capability of the network for small-scale safety helmet object features by building a context extraction module to establish multi-scale mapping of deep image features,and aggregating global contextual semantic dependencies and expanding the perceptual field.At the same time,an attention feature fusion module driven by spatial and channel attention is constructed between the underlying image features and the features sampled on the network.This module can adaptively fuse the underlying features with the upsampling features to eliminate the interference of the underlying noise on the object features,and establish the global dependency of the upsampling features to effectively enhance the network’s ability to represent the local spatial details and global association features of the object.Experimental analysis shows that the proposed method based on context extraction and attention feature fusion for safety helmet object detection has good detection performance for safety helmet objects in surveillance scenes,and the m AP can reach 86.31% in the constructed safety helmet wearing dataset,and has good detection speed.2.Safety helmet detection method based on multi-scale feature-enhanced pyramid networkTo address the problem of semantic differences caused by the direct fusion of different scale features by the feature pyramid network architecture,which results in reduced learning ability for small-scale safety helmet object features.In this paper,we propose a safety helmet detection method for multi-scale feature-enhanced pyramid networks.The method uses the constructed feature-enhanced pyramid network structure to interactively fuse shallow fine-grained features with deep coarse-grained features to generate rich contextual feature information,thereby reducing the semantic discrepancies existing between features of different scales.At the same time,a multi-scale attention module is constructed to filter out noise and redundant information between features at different scales,and a multi-scale feature progressive mapping and global semantic dependencies are established to improve the feature fusion effect in the feature-enhanced pyramid network structure and further improve the detection accuracy for small-scale safety helmet objects under surveillance images.Experimental analysis shows that the m AP of the proposed multi-scale feature-enhanced pyramid network-based safety helmet detection method can reach up to 91.95%,which has good detection effect for small-scale safety helmet targets in surveillance images.