翻译以下内容内容为英文要求英语要地道、学术化。为了克服人工计数的低效、精准性不足并实现智能地监测畜禽及相关产品数量近年来许多研究人员已经引入了适用于不同场景的视觉技术15-18。这些视觉技术可以分为两大类：1基于DIP数字图像处理的方法该方法通过阈值分割和形态学处理聚焦于目标像素特征的分析具有操作直观、易于理解的优点19-22。然而它包含了太多人工地设定参数的步骤这可能限制了它在动态场景中的应用。

In recent years, many researchers have introduced visual technologies applicable to different scenarios to overcome the inefficiency and lack of accuracy of manual counting and achieve intelligent monitoring of livestock and related product quantities [15]-[18]. These visual technologies can be divided into two categories: 1) methods based on digital image processing (DIP), which focus on the analysis of features of target pixels through threshold segmentation and morphological processing, and have the advantages of intuitive operation and easy understanding [19]-[22]. However, it involves too many steps of artificially setting parameters, which may limit its application in dynamic scenes. 2) Deep learning-driven computer vision methods, which are usually composed of deep convolutional neural networks to achieve efficient and robust feature representation and object counting in complex environments [23]-[30]. This method is known to be the most successful in commercial products, such as the pig house inspection counting robot based on key point detection (JD Finance America Corporation) [27]. However, compared with animal individuals, the smaller volume and higher density distribution of eggs make detection and counting more difficult [38]. In recent years, target detection technology [32]-[35] has matured enough to enter practical applications [36]-[38]. Ulaszewski verified the performance of three target detectors based on SSD (https://github.com/weiliu89/caffe/tree/SSD), FR-CNN, and yolo v3 (https://github.com/ultralytics/yolov3) and distance measurement for egg counting in 2021 [39], proving the feasibility of target detection in egg counting. Even with the introduction of target detection technology, it is still difficult to efficiently count eggs on the spot in the chicken house. The reason is that real breeding farms contain many interfering factors, such as high density, shade, or near-field color interference, and target detection lacks correlation with temporal information, resulting in frequent disappearance of egg bounding boxes. In addition, the limited space in the chicken house site is not suitable for deploying the model on large servers. Therefore, it is necessary to develop a reliable and lightweight egg counting algorithm applicable to edge devices for deployment in the chicken house site.