An Improved YOLOv8 Algorithm for dolomite ore Recognition

Dolomite ore as a metallurgy, building materials, refractory and other production of important raw materials, its efficient and accurate identification will help the rational use of resources. the current traditional identification methods mainly include manual inspection, chemical analysis and laser spectroscopy. Manual inspection of the strong personal subjective factors, chemical analysis requires chemical detection, and detection cycle is long, and the laser spectroscopy requires a special instrumentation and equipment, high cost. In recent years, deep learning has gradually emerged, which has a good recognition effect on the features of the image, So, in this thesis, we choose YOLOv8 model to realize the intelligent recognition of dolomite ore image, optimize the feature extraction and feature fusion ability of the network model according to the features of dolomite ore image. For the problems of leakage, misdetection and multi-detection of the model in dolomite ore recognition, the deepening of the backbone network structure is used to optimise the feature extraction structure, and the HGStem (Hourglass Stem) and HGBlock (Hourglass Block) are used as the initial core feature extraction units of the network, The dynamic convolution optimisation of the convolutional layer of HGBlock improves the ability of local feature expression and global semantic fusion, and the addition of GLSA (Global-to-Local Spatial Aggregation) achieves better delivery of global semantic information and local detailed features to the neck network. The Mean Average Precision (mAP) of the above improved model is 96.4%.