The application of object detection technology is a key component in the implementation of artificial intelligence in manufacturing and transportation. In order to achieve the classification and orientation of feature-similar objects, an enhanced method the Fine-grained Mask R-CNN (FGM R-CNN) for object detection based on self-supervised learning and fine-grained enhancement is proposed. Based on the Mask R-CNN network, the self-supervised learning for tiny feature extraction and data enhancement by cropping segmentation of individual targets are introduced. The introduced methods ensure that the network has a superior recognition ability for objects with small samples, and also enables the model to focus on the main feature regions when performing tiny feature extraction. Following that, the average accuracy of the FGM R-CNN algorithm is calculated on COCO2017 dataset. Furth more, the improved network is evaluated on the autonomously labelled dataset, and the results show that the FGM R-CNN has 36.05%, 54.54% and 38.78% in \(b\_mAP\) , \(b\_mAP\) (50) and \(b\_mAP\) (75)evaluation indexes respectively, which are improved by 1.65%, 1.19% and 1.70% compared with the Mask R-CNN. In summary, the method proposed in this paper performs superior fine-grained item classification ability for object detection, as well as has good application scenarios for similar packaging products grasping prediction.