MFSnet: a multi-scale feature screening network for chicken counting in dense environments

1. Machine-vision-based chicken counting is a highly efficient approach. Nonetheless, in scenarios with high breeding densities, chickens in the captured images frequently overlap with one another. This research addressed the challenge of accurately counting chickens within a free-range chicken coop in densely environments. It proposes a chicken-counting network specifically designed for dense scenarios, namely MFSnet.2. The study extracted multi-scale feature maps and subjected them to processing during the fusion stage via a Feature Screening Module (FSM). This module generated feature maps that were richly endowed with features from diverse scales to enhance information, thereby augmenting the network's capacity to accurately identify chickens.3. The dataset was collected and labelled and denominated as Chicken2023. It consisted of 550 images, which, in aggregate, encompassed a total of 49 747 chickens. To validate its efficacy, it was compared with extant counting algorithms. The experimental findings derived from the Chicken2023 dataset illustrated that this method attained a better counting performance level. It achieved a mean absolute error (MAE) of 2.7 and a root mean square error (RMSE) of 3.6. When juxtaposed with the top-performing network, it showed a notable improvement, with a 6.25% reduction in MAE and a 6.26% reduction in RMSE.4. The network model proposed in this study accurately recognised the number of chickens in dense environments and improved the efficiency of poultry farming.