Object detection is an important basis for understanding the high-level semantic information of images. To address the problems of small object accuracy and inaccurate bounding box localization in the YOLOv3, we propose a new object detection method, YOLOv3-XH. In YOLOv3-XH, we make three improvements to YOLOv3. Firstly, we propose the Enhanced Feature Pyramid Network (EFPN), which fuses Path Aggregation Network (PAN) and Feature Pyramid Network (FPN) and helps extract richer feature information and increase the accuracy of small object detection. Secondly, the Multi-Branch Receptive Field (MBRF) is proposed, which adds a branch to RFB for feature extraction and helps to increase the receptive field and enhance feature extraction. Finally, a new loss function is proposed, using GIo U loss as the bounding box loss function, so that the network is optimised in the direction of a higher overlap between the predicted and real boxes. Our approach achieved an average mean accuracy mAP of 83.02% (IoU=O.5) on the PASCAL VOC test set, reaching an improvement of 1.83%.