With the development of artificial intelligence, automatically and accurately counting chickens has become a reality. However, insufficient lighting, irregular sizes, and dense flocks make this a challenging task. The existing methods cannot perform accurate and stable counting. In this article, a dense flock of chickens counting network (DFCCNet) is proposed based on density map regression, where features from different levels are merged using feature fusion to obtain more information for distinguishing chickens from the background, resulting in more stable counting results. Multi-scaling is used to detect and count chickens at various scales, which can improve the counting accuracy and ensure stable performance for chickens of different sizes. Feature convolution kernels are adopted to convolve feature maps, which can extract more accurate target information, reduce the impact of occlusion, and achieve more reliable and precise results. A dataset of dense flocks of chickens (namely Dense-Chicken) has been collected and constructed, which contains 600 images of 99,916 chickens, with labeled points and boxes. It can be accessed by researchers as benchmark data. The proposed method was compared with some state-of-the-art algorithms, to validate its effectiveness. With its robustness being verified by counting in three kinds of density situations, with the mean absolute error being 4.26, 9.85, and 19.17, respectively, and a speed of 16.15 FPS. DFCCNet provides an automatic and fast approach to counting chickens in a dense farming environment. It can be easily embedded into handheld devices for application in agricultural engineering.

Accurate monitoring of biological density is crucial for optimal breeding in aquaculture. Existing counting methods struggle to maintain precision across diverse density scenarios, ranging from sparse to dense populations. This study proposes an integrative counting framework employing a multi-task strategy tailored for Chinese soft-shelled turtles. An improved YOLOv5n is utilized for sparse scenes to enhance occlusion handling, while a novel Turtle Counting Network (TCNet) based on density map regression boosts accuracy in dense settings. Densely Connected Convolutional Networks (DenseNet) precede these counting modules, classifying image density to dynamically select the appropriate model. Experiments reveal counting accuracies of 95.04%, 96.19%, and 96.24% for low, middle, and high densities, respectively. This framework outperforms state-of-the-art methods in both sparse and dense environments, providing technological support for efficient counting under irregular densities in aquaculture.