基于NAF-YOLO v8的集成运动模糊图像复原的奶牛体况评分方法

Method for Dairy Cow Body Condition Scoring Based on Integrated Motion-blurred Image Restoration Using NAF-YOLO v8

摘要:

为解决因运动导致图像模糊难以有效进行奶牛体况评分的问题，改进噪声自适应流网络（Noise adaptive flow network，NAFNet）模型对奶牛运动模糊图像进行复原，提出集成运动模糊图像复原的奶牛体况评分方法NAF-YOLO v8。将NAFNet中通道注意力（Simple channel attention，SCA）模块更换为通道先验卷积注意力（Channel prior convolutional attention，CPCA）模块，在通道和空间维度上动态分配注意力权重，完成特征细化与增强，生成高质量奶牛体况特征图;将反向残差移动模块（Inverted residual mobile block，iRMB）嵌入YOLO v8模型C2f模块中，有助于捕获奶牛体况特征在空间上的联动关系，且保持了模型轻量级;将可分离大核注意力（Large separable kernel attention，LSKA）模块嵌入到空间金字塔池化融合（Spatial pyramid pooling fusion，SPPF）模块中，增加有效感受野，增强模型对全局特征的提取能力，实现奶牛体况评分。实验结果显示，NAF-YOLO v8在奶牛运动模糊测试集上精确率、召回率和平均精度分别为80.7%、75.3%和80.1%，相较于未进行图像复原模型，分别提高8.2、4.7、8.0个百分点，表明提出的NAF-YOLO v8方法有效减小了运动模糊对体况评分准确性的影响，降低了误检和漏检概率，提高了检测精度，为奶牛智能化管理提供了重要支持。

Abstract:

Aiming to address the challenge of accurately assessing the body condition of dairy cows in environments with motion blur, an integrated method for body condition scoring that incorporated an improved noise adaptive flow network model was proposed for the restoration of motion-blurred images. The simple channel attention module in NAFNet was replaced with a channel prior convolutional attention module, which dynamically allocated attention weights in both channel and spatial dimensions, refining and enhancing features to generate high-quality body condition feature maps. Additionally, the inverted residual mobile block module was embedded within the C2f module of the YOLO v8 model, maintaining the model's lightweight nature while aiding in the extraction of spatial relationships among body condition features. Furthermore, the large separable kernel attention module was integrated into the spatial pyramid pooling fusion (SPPF) module to achieve a larger effective receptive field, enhancing the model's ability to extract global features for body condition scoring. It can effectively reduce the probability of false detection and missing detection, and improve the detection accuracy. The proposed NAF-YOLO v8 method achieved precision, recall, and mean average precision of 80.7%, 75.3%, and 80.1%, respectively, on the motion-blurred test set, representing improvements of 8.2, 4.7, and 8.0 percentage points compared with models without image restoration. This integrated approach effectively reduced the impact of motion blur on the accuracy of body condition scoring, providing significant support for the intelligent management of dairy cattle.