Embedded devices in substations frequently encounter challenges related to real-time performance and detection accuracy, especially in scenarios involving small targets and densely arranged pointer instruments. This paper proposes an enhanced substation pointer instrument recognition model based on YOLOv5s-BCGS, which improves detection accuracy and efficiency. The model employs YOLOv5s as its backbone network, incorporating a coordinate attention mechanism at the neck to enhance spatial feature extraction. Additionally, the original path aggregation network is replaced with a weighted bidirectional feature pyramid network to better integrate positional and detailed information from the feature maps. This modification increases the model’s sensitivity to target location and size, particularly in complex scenarios. To accelerate inference speed and reduce model size, we substitute traditional convolutions with lightweight Ghost Convolutions. Furthermore, the conventional Complete Intersection over Union loss function is replaced by the SCYLLA-Intersection over Union loss function, which improves both the training speed and the inference accuracy for small targets at greater distances. Experimental results show that the proposed model outperforms YOLOv5s on a custombuilt substation pointer instrument dataset, with mAP0.5 increasing by 2.2%, mAP0.75 improving by 3.8%, and mAP0.5~0.95 rising by 6.7%. Additionally, the model size is reduced by 34.07%. When compared to other widely used models such as Faster R-CNN, YOLOv4-tiny, YOLOv7-tiny, and YOLOv8n, our model shows significant improvements in both accuracy and speed. The model, with a size of only 18.0 MB, demonstrates strong generalization and robustness, making it well-suited for lightweight deployment. Inference speeds on a PC and the Jetson Xavier NX development board reach 154.7 FPS and 18.7 FPS, respectively, meeting the performance requirements for embedded devices used in substation pointer instrument inspections.