A Method for Enhancing Image Samples of Farmed Animals Based on SAM Image Segmentation

The detection and recognition of farm animals using computer vision-related technologies is an important part of achieving smart agriculture. However, acquiring and annotating a large number of farm animal images is a very challenging task. Therefore, this paper proposes a sample augmentation method based on image segmentation. First, a small dataset of real farm scene images is collected using drones, and the Segment Anything Model (SAM) image segmentation technique is employed to extract source target images of farm animals from these images. The source target images are then subjected to conventional augmentation techniques such as flipping, scaling, and adding noise. Next, the augmented source target images are randomly pasted onto new farm background images to generate new image samples. The augmented and original image datasets were trained and tested on YOLOv5s and YOLOv7 models. The datasets enhanced using the proposed method achieved mAP scores of 87.6 % and 88.8 %, respectively. Compared to the datasets without augmentation, the mAP improved by 1 % and 1.7 %, respectively. The results indicate that the proposed augmentation method effectively enhances the diversity of the dataset, and the trained models exhibit stronger robustness.