TY  - EJOU
AU  - Zhu, Linkai 
AU  - Li, Chaofan 
AU  - Yang, Lu 
AU  - Meng, Changpu 
AU  - Liu, Lu 

TI  - Towards Training-Free Long-Form Image Captioning: A Divide-and-Conquer Strategy
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Long-form image captions are valuable for image–text retrieval, visual question answering, and generative applications, but multimodal large language models (MLLMs) still face a persistent trade-off between fine-grained semantic coverage and visual faithfulness. To address this problem, we propose a training-free divide-and-conquer framework for reliable long-form image captioning. Inspired by Feature Integration Theory, the framework decomposes an image into four object-aware spatial regions and one key semantic region, samples multiple regional captions, and progressively integrates verified evidence through same-region semantic fusion, cross-region semantic fusion, and key-semantic-region-guided global refinement. To reduce hallucination during aggregation, unreliable regional and cross-region candidates are removed by image–text consistency verification, and the final caption is further filtered at the semantic-unit level by combining Contrastive Language-Image Pre-training (CLIP) alignment, Bootstrapping Language-Image Pre-training with frozen image encoders and large language models (BLIP-2) generation consistency, and hyperspherical feature similarity. Experiments on the Detailed Image Description Benchmark (DID-Bench) across open-source Large Language and Vision Assistant (LLaVA) models and proprietary MLLMs show that the proposed framework improves caption quality, semantic coverage, visual consistency, and hallucination suppression without retraining or fine-tuning. In particular, it achieves fine-grained semantic-coverage scores of 58.88 and 64.48 on LLaVA-1.5-7B and LLaVA-1.6-7B, respectively. On the 400-image ImageInWords subset (IIW-400), the framework obtains the best Bilingual Evaluation Understudy (BLEU) and Semantic Propositional Image Caption Evaluation (SPICE) scores among the evaluated methods, although other automatic metrics remain model- and metric-dependent. A human evaluation on 50 randomly selected IIW-400 images further shows the highest factual-correctness and detail-coverage scores, while the Image Textualization (IT) baseline retains the highest readability score. These results indicate that the proposed framework can improve visually grounded long-form captioning in the evaluated natural-image settings, with remaining limitations in inference efficiency and domain generalization.
KW  - Image captioning; multimodal large language models; semantic fusion; hallucination mitigation; training-free framework

DO  - 10.32604/cmc.2026.087313