【摘要】 针对复杂工况环境下矿石分割精度低的问题，提出改进Mask RCNN实例分割算法。Mask RCNN采用简单的反卷积操作来恢复矿石的掩膜，导致矿石边缘信息丢失，因此提出混合注意力模块，该模块将通道注意力机制和空间注意力机制进行加权融合，可以在通道层面和空间层面对矿石的特征信息加强。结合混合注意力模块改进Mask RCNN的分割网络结构，减少在卷积运算中造成的矿石细节信息丢失，提高网络对矿石的分割精度。用制作的矿石数据集进行网络训练和网络评价，将改进之后的算法与原算法以及其他两种常用算法进行对比试验研究，试验结果表明，改进算法中MIoU约为87.1%,相较原算法提升约19.15%,MPA值约为94.61%,较原算法提升了约22.93%,对于矿石识别分割具有更高的检测精度。改进算法在复杂工况环境下对矿石的精确识别和分割有很大的应用前景。更多还原

【Abstract】 Aiming at the low precision of ore segmentation under complex working conditions, an improved Mask RCNN instance segmentation algorithm was proposed. Mask RCNN used a simple deconvolution operation to restore the mask of the ore, resulting in the loss of edge information of the ore. Therefore, a mixed attention module was proposed, which combined the channel attention mechanism and the spatial attention mechanism, and can strengthen the feature information of the ore at the channel level and the spatial level. Combined with the mixed attention module, the segmentation network structure of Mask RCNN was improved to reduce the loss of ore details in the convolution operation and improve the segmentation accuracy of the network for ore. The network training and network evaluation were carried out with the ore data set, and the improved algorithm was compared with the original algorithm and the other two common algorithms. Experimental results show that the improved algorithm MIoU is about 87.1%, which is about 19.15% higher than the original algorithm. MPA value is about 94.61%, which is about 22.93% higher than the original algorithm. It has higher detection accuracy for ore identification and segmentation. The improved algorithm has a great application prospect for accurate identification and segmentation of ore under complex working conditions.更多还原