The setting and operation of beneficiation process parameters for different ore types are conducive to ensuring the quality of concentrate, improving the recovery rate and reducing the material consumption. Due to the lack of effective means for real - time detection of ore types, it is difficult to achieve accurate operation in the beneficiation process. Therefore, a comprehensive improved Mask R - CNN ore type detection algorithm is proposed. Algorithm steps: 1) ResNetV1d50 is used to extract the feature maps of each stage of the ore image, and deformable convolutions are added to the backbone network to enhance the characteristics of the irregular - shaped ore; 2) To improve FPN, a multi - scale feature map with stronger semantic information is obtained by adding a feature residual module to the C5 feature layer of the backbone network and fusing it into the P5 feature layer; 3) Improve the RPN, and design an adaptive positive sample IOU selection scheme to match the ore with abnormal aspect ratio, so as to further improve the identification accuracy of special - shaped ore; 4) Global - Context is added on the basis of the RoIAlign network to improve the detection ability of small ores; 5) The model is improved in terms of data augmentation and training skills. The results show that the average accuracy of this algorithm is 67.92%, and the average intersection ratio is 63.54%, which is 13.67% and 9.71% higher than the benchmark model. The research method has good application value in ore type identification.