Fonisto, Mattia (2025) Contactless Mastitis Screening in Automated Milking: A Radiometry-to-Decision Pipeline with Self-Supervised Computer Vision at the Edge. [Tesi di dottorato]

Documento PDF mattia_fonisto_38_COMPLETO.pdf Download (11MB)

Collection description

Subclinical mastitis (SCM) in Mediterranean buffalo reduces milk quality and yield while remaining largely invisible to routine inspection. This thesis develops a pipeline that couples radiometric infrared thermography (IRT) with self-supervised computer vision and low-power edge deployment to enable contactless first-stage mastitis screening integrated with automated milking. The work addresses five coupled challenges: (i) reproducible, robot-timed thermal acquisition with ambient compensation; (ii) pixel-accurate udder segmentation under barn conditions and scarce labels; (iii) principled conversion from radiance to udder skin surface temperature (USST) with atmospheric modeling; (iv) clinical linkage of USST indicators to reference markers (SCC/DSCC and on-line cell count); and (v) an on-premises architecture that privately serves multiple barns with constrained compute. A curated, expert-annotated buffalo-udder thermography dataset (2,148 image–mask pairs) is released to standardize benchmarking. Transformer-based segmentation (SegFormer, MiT-b5) establishes strong supervised baselines on the held-out test set (mIoU = 0.827; detection accuracy = 77.9% at 0.80 IoU). Self-supervised pretraining on unlabeled thermograms (DenseCL first, DINOv2 last) improves the encoder's ImageNet initialization, supporting label efficiency when annotation budgets are limited. A radiometry-to-alert pipeline is formulated that inverts atmospheric effects, derives animal-normalized USST statistics within model masks, and issues milking-time screening alerts to prompt further veterinary investigation. In a prospective in-barn study, the maximum USST is significantly higher in high-OCC milkings (threshold 400,000 cells/mL), with a mean shift of 0.71°C (Welch’s t-test p=0.002) after compensation. Leveraging this signal, a temporal screening model for early-stage mastitis (threshold 200,000 cells/mL) achieved a recall of 86%, effectively filtering the herd for targeted inspection despite data scarcity. The mean absolute discrepancy between Tmax computed on expert versus model masks is 0.159°C, indicating limited measurement bias for downstream decisions. For deployment, a pruned and quantized ConvNeXtTiny+FPN initialised via DINOv3 weights model achieves near-baseline accuracy (mIoU = 0.819) with static INT8 post-training quantization and runs at 165.1ms per frame on a Jetson Orin Nano (TensorRT EP, 7W), enabling private, on-prem multi-barn servicing. Overall, this thesis work demonstrates that radiometrically grounded IRT, label-efficient segmentation, and edge AI can deliver an actionable, privacy-preserving screening tool for buffalo mastitis management in real farm settings.

Downloads

Actions (login required)

View Item