Abstract: This paper presents a method for estimating the number of grassland sheep based on an improved CSRNet with a multiscale residual feature fusion network. The approach addresses the challenges of low counting accuracy in the livestock industry caused by large scale variations and severe occlusion of sheep in pastures, which traditional counting algorithms struggle to handle effectively. To enhance performance, the method improves upon CSRNet by incorporating dense dilated convolutions to construct dense multiscale residual modules, which are integrated into the backbone network to extract multiscale features of sheep targets, making the model more adaptable tovarying scales. Additionally, a multibranch feature extraction network was constructed to optimize the output feature maps of the backbone network, thereby enhancing the overall feature extraction capability and improving counting accuracy. Furthermore, the CBAM attention module was incorporated to strengthen the models ability to capture sheep position features and further refine the output density map. Experimental results demonstrate that the proposed method achieved MAE and RMSE values of 12.3 and 13.9, respectively, significantly outperforming the five advanced counting methods: CSRNet, MCNN, DSNet, PaDNet and HA—CCN. The model exhibited high robustness and accuracy in scenarios with dense sheep distribution, severe occlusion, and lowlight conditions, hence showcasing its superior performance in grassland sheep counting tasks.

Key words:  , sheep counting, convolutional neural network, multiscale, dilated convolution, deep estimation

摘要： 针对牧场内绵羊计数中目标尺度变化大、遮挡严重等导致的漏检误检、计数精度低的问题，提出一种基于改进CSRNet的多尺度残差特征融合网络的草原绵羊数估计方法。该方法在CSRNet基础上进行改进，利用密集扩张卷积构建密集多尺度残差模块，嵌入到模型骨干网络中，用于提取绵羊目标的多尺度特征，更好地适应绵羊的多尺度变化。此外，构建多分支特征提取网络，优化骨干网络输出特征图信息，加强模型整体特征提取能力，进而提高计数精度。同时，引入CBAM注意力模块，加强绵羊位置特征的表达能力，进一步修正输出密度图。试验结果表明，所提方法的平均绝对误差MAE和均方根误差RMSE分别为12.3、13.9，明显优于CSRNet、MCNN、DSNet、PaDNet和HA—CCN五种主流计数方法；且在羊群密集分布、遮挡严重和光照不足的情况下展现出较高的鲁棒性和准确性，证明其在草原羊群计数任务中的优越性能。

关键词: 绵羊计数, 卷积神经网络, 多尺度, 扩张卷积, 深度估计