Abstract:

To address the issues of poor image quality， difficulty in recognition and localization， as well as false positives and false negatives caused by occlusion and adhesion in scenarios involving clustered black pigs under low-light conditions， a detection model named low-light animal detection network （LADnet） was proposed. Firstly， an illumination-adaptive transformer （IAT） and a coordinate attention （CA） mechanism were utilized to enhance the brightness and reduce noise in the images. Then， a selective kernel convolutional attention （SKCA） module was designed to improve the model’s ability to perceive black pigs. Finally， the ReLU activation function was employed to mitigate problems related to gradient vanishing and explosion. The results showed that the LADnet model achieved precision， recall and mean average precision （mAP@0.5） of 97.32%， 86.61% and 92.73%， respectively， representing improvements of 1.07， 6.15 and 3.05 percentage points compared to the baseline model. Compared to single-stage object detection models such as SSD and YOLOv5， LADnet achieved an average accuracy improvement of 8.33 and 7.35 percentage points， respectively. In comparison with two-stage models like Cascade R-CNN， Faster R-CNN and DAB_DETR， LADnet not only demonstrated higher detection accuracy but also achieved a smaller parameter size and faster detection speed， making it more suitable for the real-time detection requirements. The LADnet model demonstrated exceptional detection performance and enhanced robustness in low-light black pig detection tasks， providing an efficient and reliable tool for the accurate identification of black pigs in low-light environments， which holded significant importance for advancing the development of intelligent farming under low-light condition.

Key words: black pig inventory, object detection, attention mechanism, low-light enhancement, feature extraction, YOLOv7, intelligent breeding

摘要：

为解决低光环境下黑猪图像质量差、识别与定位难度大，以及黑猪聚集时因遮挡和粘连导致的误检与漏检问题，提出LADnet（low-light animal detection network）检测模型。首先，采用亮度自适应模块（illumination-adaptive-transformer，IAT）及坐标注意力机制 （coordinate attention，CA）对图像进行增亮与降噪处理；接着，设计了选择核卷积注意力（selective kernel convolutional attention，SKCA）模块，以提高模型对黑猪的感知能力；最后，利用ReLU激活函数解决了梯度消失和梯度爆炸等问题。结果表明，LADnet模型的精确率、召回率和平均准确率（mAP@0.5）分别达到97.32%、86.61%和92.73%，与基准模型相比，分别提升1.07、6.15和3.05个百分点；与单阶段目标检测模型SSD、YOLOv5相比，LADnet的平均准确率分别高出8.33、7.35个百分点；与两阶段目标检测模型Cascade R-CNN、Faster R-CNN、DAB_DETR相比，LADnet在检测精度提高的同时，具有更小的参数量和更快的检测速度，更符合实时检测要求。LADnet模型在低光黑猪检测任务中展现出卓越的检测性能和更高的鲁棒性，为低光环境下的黑猪精准识别提供了一种高效可靠的工具，对推动低光智慧养殖的发展具有重要意义。

关键词: 黑猪盘点, 目标检测, 注意力机制, 低光增强, 特征提取, YOLOv7, 智慧养殖