Abstract: Aiming at the problems of low detection accuracy, high missed detection rate, poor adaptability to complex scenarios and others in coal gangue target identification, a DCB-YOLO intelligent identification model (The model name is composed of the acronyms of the three core improvement modules) that integrates deformable convolution (DCNv2), coordinate attention (CA), and bidirectional feature pyramid (Bi FPN) is proposed. The model improves performance through three items of key improvements: embedding deformable convolutions in the backbone network C3K2 module to construct C3K2_DCNv2 module, enhancing the feature extraction ability for irregular coal gangue targets; Introducing coordinate attention mechanism to enhance the model^,s attention rate to the spatial position information of coal gangue targets; Adopting a weighted bidirectional feature pyramid to optimize the fusion efficiency of multi-scale features and improve the identification accuracy under small targets and complex backgrounds. The experimental results show that the detection accuracy, accuracy rate, recall rate and real-time detection speed of the DCB-YOLO model reach 98.4%, 95.8%, 95.6%, and 103.1 FPS, respectively, compared to YOLOv11n, it is improved by 1.3%, 0.6%, 2.1%, and 2.0 FPS, respectively, and the parameter count is reduced from 2.6 × 10⁶ to 1.9 × 10⁶ (with a decrease of 25.2%). This model can achieve precise identification of coal gangue targets under complex scenarios such as noise interference, low illumination, motion blur and others, which provides an efficient technical scheme for the intelligent upgrade of coal sorting links.