Low-resource domains suffer from limited labeled data, which significantly degrades the performance of deep learning models trained from scratch. Transfer learning and fine-tuning methods have advanced as significant ways to tackle this problem by using the knowledge from those domains with a large amount of data. The present research carries out a performance and generalization comparison among the baseline convolutional neural networks, transfer learning models, and fine-tuned neural networks in a setting of low-resource domain adaptation. A statistically validated experimental framework is applied, where accuracy and F1-score are used as the evaluation metrics. The use of randomized controlled trials combined with post-hoc tests made it possible to detect very significant differences between the fine-tuned neural networks and the baseline ones (p < 0.005). This indicates that both lifting and generalizing models' data predictions to places with small datasets would be possible. The results offer practical guidance for the implementation of deep learning systems in real-world low-resource scenarios.