Research On Image Enhancement In Mines And Miner’s Helmet Detection Methods

With the development of intelligent security monitoring technology,using object detection algorithm to detect miners’ safety helmets has become an important means to ensure the safety of mining production and workers’ lives.However,the quality of the images collected in the mine is poor,showing the characteristics of low illumination,which brings difficulties to the subsequent detection work.In addition,the existing object detection algorithm is not effective in application,there is a problem of missed and false detection,there is still room for improvement and enhancement.In view of the two practical needs of the above security projects,in this thesis,the low-illumination image enhancement method in the mine and the miner’s helmet detection method was carried out.including:A low illumination image enhancement method in mine was investigated.The method flow is as follows: First,make a copy of the original image,and perform color compensation and contrast limited adaptive histogram equalization on the first image,Transfer the second image to HSV space and separate the illumination layer and reflection layer of V channel through multi-scale Gaussian function,The brightness of the illumination layer is stretched,and the new V channel is obtained after gamma correction of the reflection layer,and then the new V channel is fused with the H and S channels and inversely converted to RGB space,Finally,the two images are fused by wavelet transform to complete the whole enhancement process.The simulation experiment compared seven methods commonly used in security monitoring projects from three aspects: subjective evaluation,histogram distribution and indicators evaluation.The results showed that the enhanced images obtained by this method have a good look and feel,and the histogram distribution is relatively uniform,ranking the top in five indicators,and this method can be helpful for subsequent detection tasks.A helmet detection method based on improved YOLOv5 was investigated.Aiming at the problem of insufficient receptive fields in convolutional neural networks(CNN),a multi head self attention mechanism was added to the model to process the features extracted by CNN,making the network have the ability to capture long-distance association information between pixels similar to Transformer,and the local sensitive hash method was introduced to sparsize it,to solve the problem of too large amount of self-attention computation.In view of the problem that the preset anchor can not fit the object scale well,the K-means++method was used to cluster to get a better anchor frame.Introduced EIo U＿Loss as the location loss function.The helmet detection task was incorporated into the Pretrain-fintune framework,and the combined strategy of Noisytune and Bitfit was adopted in the training process.The results of ablation experiments showed that all improvements have a positive effect,and the improved method has a 1.98 percentage point increase in m AP compared to baseline in the dataset provided by Uniview.Comparative experimental results showed that this method outperforms classic object detection algorithms and object detection algorithms for dim light environments in recent years on both the Uniview dataset and the public dataset SHWD.A helmet detection method based on improved YOLOX was investigated.Aiming at the slow reasoning speed of complex networks with multi-branch structure,combining the idea of channel split,channel shuffle and structure re-parameterization,a lightweight module Mix-Block was constructed to replace the C3 module in the backbone to realize multi-branch training and single-way reasoning.Combining Ghost-Block with path aggregation network,Ghost-PAN was built as the new neck network.In the process of feature fusion,Ghost cheap operation was used instead of ordinary convolution to obtain more features with a smaller amount of parameters.The results of ablation experiments showed that the improved method m AP increased by 2.16 percentage points,while the weight size is reduced by 15.65%;The parameter quantity decreased by 18.19%;The computational complexity is reduced by 7.83%;The inference speed is increased by 45.57%.The comparative experimental results showed that this method performs better than the classical lightweight object detection algorithm in recent years on both the Uniview dataset and the public dataset SHWD.