Development of Attention-Enabled Multi-Scale Pyramid Network-Based Models for Body Part Segmentation of Dairy Cows

Automated assessment of dairy cow traits, important for productivity evaluation, provides advantages by mitigating personal biases, measurement errors, and stress factors typically associated with manual assessment. To develop such a system, the initial step involves accurately segmenting cow body regions for subsequent trait measurement.

Thus, the present study introduces a refined DeepLabV3 + CNN model with EfficientNetB2 as the backbone and enhanced with attention mechanisms, aiming for precise segmentation of cow body regions from lateral and posterior views. In the DeepLabV3 + model, various backbone models, including MobileNet, MobileNetV2, MobileNetV3, EfficientNetB0, EfficientNetB1, and EfficientNetB2, were evaluated. Among these, EfficientNetB2 exhibited superior performance in lateral view segmentation, achieving a mean Intersection-over-union (m-IoU) of 94.19%. To further enhance segmentation accuracy, attention mechanisms such as Squeeze and Excitation (SE), Residual connection-infused Squeeze and Excitation (SER), Convolutional Block Attention Module (CBAM), and Residual connection-infused Convolutional Block Attention Module (CBAMR) were incorporated into the DeepLabV3 + model.

The introduction of attention mechanisms in the EfficientNetB2 model led to enhanced m-IoU values: SE (94.27%), SER (94.25%), CBAM (94.59%), and CBAMR (94.66%). EfficientNetB2, integrated with CBAM and Residual connections (termed CBAMR), found to be top-performing model, achieving m-IoU values of 94.66% (lateral view), 93.77% (posterior view), and 99.61% (stature). The lateral view segmentation demonstrated high IoU for the body (98.73%) and rump (96.54%), with lowest IoU for teats (79.70%) due to their smaller spatial presence in input image. For posterior view regions, the CBAMR model achieved IoU scores above 79.0%, with the rear leg showing the highest (96.70%) and rump bones the lowest (79.52%). The segmentation accuracy for stature exceeded 90.0%, indicating less complexity in single-body region segmentation.

Therefore, these developed models demonstrate considerable accuracy in segmenting cow regions, making a significant contribution to the advancement of computer vision–based systems for measuring linear-type traits, and hold promise for deployment in such an automatic system.