<![CDATA[Meta-learning is an approach designed to address data limitations in few-shot learning scenarios. The performance of meta-learning is influenced by the quality of the initial weights used during the meta-training process. Initial weights derived from a relevant domain have the potential to produce more informative feature representations, thereby enabling the adaptation process to new tasks to proceed more effectively. This study analyzes the impact of domain-specific pretrained initialization on classification performance, learning stability, convergence behavior, and computational trade-offs within the First-Order Model-Agnostic Meta-Learning (FOMAML) framework using an enhanced ResNet-50 backbone. Experiments were conducted on a 3-way classification scenario with 1-shot, 5-shot, and 10-shot configurations. Model evaluation was performed using accuracy, precision, recall, and F1-score, while learning stability was analyzed using standard deviation (Std) and coefficient of variation (CV). The experimental results show that a chili-domain pretrained initialization consistently yields better performance than random initialization. Accuracy reached 95.33%, 95.60%, and 95.94% in the 1-shot, 5-shot, and 10-shot scenarios, respectively an increase of 17.20, 12.61, and 17.49 percentage points compared to random initialization. In terms of stability, the CV values decreased to 1.00%, 0.59%, and 1.03%, compared to 1.10%, 3.66%, and 2.42% with random initialization. These performance improvements were achieved with relatively small differences in training time 0.056 minutes, 0.170 minutes, and 0.876 minutes for the 1-shot, 5-shot, and 10-shot scenarios, respectively. Domain-specific pretrained initialization produces more relevant initial feature representations, thereby improving the effectiveness and stability of FOMAML adaptation while maintaining computational requirements comparable to those of random initialization]]>
