Automatic image captioning research is predominantly focused on resource-rich languages like English. Recent advances in large language models (LLMs) have further enhanced captioning performance for English by improving fluency and generalization. However, despite having over 15 million native speakers, the task of image captioning in Assamese remains underdeveloped. This is due to the lack of annotated datasets and pretrained language models. This paper introduces AsCap-GPT2, an Assamese image captioning model that leverages a pretrained GPT-2 model for caption generation. The proposed AsCap-GPT2 model integrates salient image regions as visual features with the pretrained GPT-2 decoder to generate captions in Assamese. The model is evaluated across three experimental settings. First, few-shot training is conducted on the COCO-AC dataset using 0.1%, 1%, 10%, and 100% of the training data. Second, the model is fine-tuned using reinforcement learning technique to directly optimize the CIDEr score. Third, its adaptability is assessed on the Flickr30K-AC dataset. On COCO-AC, AsCap-GPT2 achieves BLEU-1 of 72.5, BLEU-4 of 31.8, and CIDEr of 94.3, outperforming all baselines.