Abstract: The objects in the fully mechanized working face have the features of high-speed motion, multi-scale, occlusion, etc. The existing object detection algorithms have problems such as low precision, large memory of models, and strong hardware dependence. In order to solve the above problems, a large block coal detection algorithm based on MES-YOLOv5s is proposed in fully mechanized working face. The method adopts a lightweight design, uses MobileNetV3 as the backbone network to reduce the memory occupied by the model and improve the detection speed on the CPU side. The method adds an efficient multi-scale attention (EMA) module to the neck network, fuses contextual information of different scales, and further reduces computational overhead. The method uses SIoU loss function instead of CIoU loss function to improve training speed and inference accuracy. The ablation experiment results show that MobileNetV3 significantly reduces the memory and detection time occupied by the model, but the mAP loss is severe. The EMA module and SIoU loss function can restore the precision of the loss to a certain extent, while ensuring that the model has a high detection speed on the CPU, meeting the real-time detection needs of coal mine underground objects. The comparative experimental results show that compared with DETR, YOLOv5n, YOLOv5s, and YOLOv7 models, the MES-YOLOv5s model has the best overall performance, with an mAP of 84.6%. The model occupies 11.2 MiB of memory and has a detection time of 31.8 ms on the CPU side. It can maintain high recall and precision in high-speed motion, multi-scale, occlusion, and multi-object working environments.