基于多尺度残差视觉信息融合的牧场牛只数量估计方法

由于牧场牛只分布不均以及尺度变化大，传统的目标计数算法在畜牧领域计数精度不高，且用于研究的牛只数据集较少。针对这些问题创建了一个用于牛只密度估计的数据集，并提出了一种基于多尺度残差视觉信息融合的牧场牛只数量估计方法。该方法利用多个并列且空洞率不同的空洞卷积提取牛只目标的多尺度特征，并将残差结构与小空洞率卷积相结合，设计出更适合牛只活体计数的深度神经网络，从而缓解了由空洞卷积带来的“网格效应”的影响，同时能更好地适应牛只的多尺度变化。在牛只密度数据集中，该方法取得了最低的平均绝对误差（MAE）和均方根误差（RMSE）。此外，在密集人群数据集中，该方法的MAE和RMSE也属于最优或次优结果。实验结果表明，该方法不仅适用于牛只场景的数量估计，在人群密度估计中也有较高的准确性和鲁棒性。

Cattle counting estimation method based on multi-scale residual visual information fusion

Due to the uneven distribution and large scale-variation range of living cattle, the accuracy of traditional object counting algorithm is not high in living-animal field, and there are few cattle data sets used for research. To solve these problems, this paper established a dataset for cattle density estimation, and proposed a cattle counting estimation method based on multi-scale residual visual information fusion. This method used multiple parallel dilate convolution with different dilate rates to extract multi-scale features of cattle, and combined residual structure with small dilate rate convolution to design a deep neural network more suitable for living cattle object counting, thus eased the effect of "grid effect" caused by dilate convolution and better adapt to the multi-scale changes of cattle. The proposed method achieved the lowest mean absolute error(MAE) and root mean square error(RMSE) on the cattle density data set. In dense counting data sets, the MAE and the RMSE of the proposed method also achieved the optimal or suboptimal results. The experimental results show that the proposed method is not only suitable for the number estimation of cattle scene, but also has high accuracy and strong robustness in population density estimation.