基于YOLO v8-ST的叠层笼养肉鸡死鸡识别方法

Identification Method of Dead Broilers in Stacked Cages Based on YOLO v8-ST

摘要:

针对规模化商品肉鸡养殖场中死鸡自动检测问题,构建了一种基于改进 YOLO v8 的检测模型。 基于自建数据集,提出 YOLO v8-ST 模型,在 C2f 模块中引入移位稀疏卷积(Shift-wise Conv),通过小核卷积模拟大核感受野以增强复杂场景下的特征提取能力;同时引入 Power Transform(PT)函数优化对齐度量中的 Overlap 计算,以提升高质量预测框的学习能力。 实验结果表明,YOLO v8-ST 模型的精确率、召回率和平均精度分别达到 94. 7% 、92. 8% 和 97. 6% ,较基线分别提升 4. 2、4. 1、4. 0 个百分点。 针对模型复杂度较高的问题,引入 LAMP 剪枝方法进行轻量化优化。 在全局剪枝率为 33% 时,模型平均精度仅下降了 2. 0 个百分点,而浮点计算量降低约 33. 0% ,参数量与模型内存占用量(9. 1 MB)分别减少 60. 0% 和 57. 0% 。 多相机实验结果表明:单相机模型在同源数据上性能最优;多相机数据融合可显著提升检测性能;低照度相机在低照度环境下具有替代高价工业相机的潜力。

Abstract:

A detection model based on the improved YOLO v8 was constructed to address the issue of automatic dead chicken detection in large-scale commercial broiler farms. Based on a self-built dataset, the YOLO v8-ST model was proposed. In the C2f module, the shift-wise convolution was introduced to simulate large kernel receptive fields through small kernel convolutions, thereby enhancing the feature extraction capability in complex scenarios. Meanwhile, the Power Transform (PT )function was introduced to optimize the overlap calculation in the alignment metric, improving the learning ability of high-quality prediction boxes. Experimental results showed that the precision, recall, and average precision of the YOLO v8-ST model reached 94. 7% , 92. 8% , and 97. 6% , respectively, which were 4. 2, 4. 1, 4. 0 percentage points higher than the baseline. To address the issue of high model complexity, the LAMP pruning method was introduced for lightweight optimization. When the global pruning rate was 33% , the average precision of the model was decreased by only 2. 0 percentage points, while the floating-point operation count was reduced by approximately 33. 0% , and the number of parameters and model size (9. 1 MB)were reduced by 60. 0% and 57. 0% , respectively. The results of multi-camera experiments indicated that the single-camera model performed optimally on homologous data; multi-camera data fusion could significantly improve detection performance; and low-illumination cameras had the potential to replace expensive industrial cameras in low-illumination environments. The methods and conclusions proposed for dead chicken detection can provide theoretical references for the automatic inspection of dead chickens in broiler farms.