Abstract: In the coal gangue sorting scene, coal and gangue are closely intertwined, making it challenging for the mechanical claw to accurately grasp gangue. To enhance grasping reliability under complex conditions, a multi-objective small-gap coal gangue recognition method based on semantic segmentation is proposed. The FCN (Fully Convolutional Networks) algorithm is improved by integrating the FPN (Feature Pyramid Network) module and replacing the cross-entropy loss with Dice Loss. A total of 1 202 coal gangue images are annotated using Labelme, and enhanced with the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm. The performance of various semantic segmentation algorithms and transfer learning strategies are analyzed by comparative experiment. Results show that the FCN-ResNet50-FPN model, with ResNet50 as the feature extractor, achieved a precision of 95. 0%, a recall of 95. 4%, an F1 score of 95. 2%, and a mIoU(mean Intersection over Union) of 90. 9%. Transfer learning further improves recognition, enhances small-gap detection and provides reliable data for precise coal gangue sorting robot operations. 

Key words: multi-objective, coal-gangue identification, semantic segmentation, transfer learning