Mechanisms And Methods For Blast-hole Detection Application Based On Semantic Segmentation With Semi-supervised Learning

Blast-hole detection is an essential part of explosive loading.There are problems of low efficiency,high cost,and easy injury in traditional manual detection.The realisation of automatic blast hole detection is important in safe production applications.In this paper,the following research work was carried out with the objective of high accuracy blast-hole detection.Based on deep learning and fully convolutional neural networks,an ERF-AC-PSPNet model with symmetric Encoder-Decoder structure is proposed.The causes of losses in model feature extraction are analyzed,and a feature extraction strategy using an efficient residual decomposition network fused with an attention mechanism is proposed for the model input to obtain a feature map with fewer feature losses,which provides a basis for segmentation network design.The global average pooling and horizontal global prior methods are used to analyze the contextual information between different sub-regions in the feature map and propose a segmentation network implemented based on the feature map fusing multi-scale information and global feature vectors.The model is validated on the blast-hole segmentation task,and the Io U and Dice reach 0.945 and0.981 respectively.A fused semi-supervised learning algorithm based on the self-training algorithm and the pseudo-labelling algorithm in semi-supervised learning is proposed to address the problem that the labeled samples of the blastholes are few and the unlabeled samples are generally effective in training the model.The method extends the core idea of semi-supervised learning to semantic segmentation,designs a simple self-training algorithm to analyze the factors affecting the training of models with unlabeled samples,and proposes an improved self-training algorithm based on priority scheduling to reduce the overfitting in training.By comparing the labeling generated by the self-training algorithm with the ground truth and errors,the pseudo-labeling algorithm is introduced to address the limitations of the impact of self-training labeling on the model,and a classifier based on the HOG algorithm and deep data is designed to enhance the impact of complex samples on the model in the pseudo-labeling algorithm without increasing the complexity of the model.The experimental results demonstrate that the semi-supervised learning algorithm is effective in reducing the dependence of the model on labeled data and enhancing the impact of unlabeled data on the training of model parameters,laying the foundation for the practical application of the method.17 Figures,13 Tables and 63 References...