Bovine mastitis, a prevalent and economically detrimental inflammatory disease in dairy cattle, leads to substantial economic losses due to reduced milk yield, impaired milk quality, and increased veterinary costs. Infrared thermography (IRT) provides an efficient, non-invasive approach for early mastitis detection by leveraging computer vision to analyze temperature differences between ocular and udder regions. However, in practical farm environments, arbitrary head movements, particularly horizontal rotation, alter sensor-to-object distance, introducing systematic biases in thermal measurements and degrading diagnostic reliability. To address this limitation, a two-stage detection framework is proposed under a lateral infrared imaging configuration, in which the camera is positioned perpendicular to the walking direction of cows. First, a posture classification module based on pose estimation is employed to identify cows with consistent lateral orientation and minimal head rotation, thereby maintaining stable sensor-to-object distance between the ocular and udder regions. Second, instance segmentation is applied to posture-constrained thermal images to achieve accurate regional temperature extraction and quantitative analysis. In addition, an improved YOLO26 model integrated with a Deformable Attention Transformer (DAT-YOLO26) is developed to enhance feature representation capability and detection performance under variable imaging conditions. On an independent 40-cow blind diagnostic cohort (20 healthy, 20 mastitis-positive), the proposed method achieved an accuracy of 87.5%, an F1-score of 87.18%, a sensitivity of 85%, and a specificity of 90%. These results, while constrained by the size of the test cohort, suggest that enforcing posture consistency improves IRT-based diagnostic performance under the present acquisition conditions. We emphasize, however, that these findings are based on a single-farm, single-season pilot cohort of 40 cows; multi-farm and multi-season validation are required before broader deployment claims can be made.