The Research Of Helmet Wearing Detection Based On Convolutional Neural Network

It is significant for safety production to the research of helmet wearing detection.The target positioning of staff based on helmet detection is a typical application of computer vision in the field of target detection.In recent years,domestic and foreign research scholars have proposed many target detection algorithms and conducted a large number of experiments.The Single Shot Multi-Box Detector(SSD)algorithm is one of the commonly used algorithms based on image background target detection.The SSD algorithm is based on a feed forward neural network,which transforms the target frame positioning problem into a regression problem.As well as using the idea of regression and the candidate frame mechanism,a single deep convolutional neural network is used for target detection tasks.This paper mainly introduces the research of original SSD algorithm and the research of improved SSD algorithm for the detection of helmet wearing.The research work of the thesis is as follows:1.Build a helmet detection dataset based on helmet wearing detection,expand the data set by rotating and zooming the image,and adjust the basic parameters of the original SSD algorithm on the helmet detection dataset,train the model,and test detection in order to completes the preliminary study of the original SSD algorithm.2.Since the original SSD algorithm cannot effectively detect the helmet,for example,in the helmet dataset,due to the influence of factors such as the small target of the helmet and the occlusion of the target,an improved SSD algorithm incorporating self-attention is proposed.The force mechanism expands the receptive field of the SSD algorithm and enhances the recognition ability of the algorithm in order to cope with small target problems,and the feasibility of the algorithm is verified by the test set of the helmet data set.3.Due to the network in the original SSD algorithm does not fully utilize the shallow features extracted by the network.Use the idea of multi-scale features to build a feature pyramid and integrate it into the original SSD algorithm.Meanwhile,add deformable convolution and expansion convolution to the SSD algorithm,in order to deal with the problems of small target detection and lack of field perception,and through comparative experiments It is verified that the improved algorithm’s ability to perceive small targets is improved compared to the original algorithm.