A lightweight model for automatic pig counting in intensive piggeries using a green inspection robot and image segmentation method

Abstract

To address the high computational resource consumption of traditional pig instance segmentation models, which impedes their deployment on resource-constrained edge devices, this paper proposes an improved, lightweight instance segmentation and counting method based on YOLOv8n-seg model. Specifically, the C2f module is replaced with the Ghost module to reduce the model’s computational complexity. Additionally, a spatial group-enhanced attention mechanism is introduced in the neck network to enhance the model's feature fusion ability in the presence of pig occlusion and overlap. In the head network, a lightweight shared detail-enhanced convolution detection head is employed, which reduces computational load and parameter count through shared convolutions while capturing the intricate details of pigs from multiple angles via the detail-enhanced convolution module. Experimental results show that the improved model achieves an average precision of 95.7 % with memory usage, floating-point operations per second (FLOPS), and frames per second (FPS) at 1.2 MB, 7 × 10^9, and 217.86, respectively. Compared with State-of-the-art model, such as DeeplabV3+, HRNet, PSPNet, Seg-Former, and UNet models, the proposed model exhibits superior performance metrics. This research provides a lightweight solution for pig instance segmentation in farm environments.