Deep learning for fine-grained image classification typically demands both extensive labeled data and expensive full-model fine-tuning. In this work, we present MeowWoofNet, a hybrid framework that combines FixMatch semi-supervised learning (SSL) with Gradient-based Parameter Selection (GPS) to address these bottlenecks in the Oxford-IIIT Pet Dataset. The FixMatch approach is employed to leverage unlabeled data in scenarios with limited labeled supervision, demonstrating that comparable 91.74% accuracy to fully supervised models can be achieved with as little as 50% labeled data. A systematic analysis of pseudo-label confidence thresholds reveals important trade-offs between label quality and supervisory signal quantity. Furthermore, GPS-based masked fine-tuning is introduced to significantly reduce the number of trainable parameters, yielding competitive performance while improving computational efficiency. Finally, integrating GPS into the FixMatch loop yields $(92.19 \pm 0.11)$ % accuracy---only 0.73 percentage points below GPS alone---validating compatibility. Our results indicate that MeowWoofNet attains near–state-of-the-art performance with minimal annotation and compute, offering a practical path for resource-efficient transfer learning.

Code base: https://github.com/Sebelino/DD2424-project