RCVQA: Visual question answering model based on reading comprehension

Visual Question Answering (VQA) combines computer vision and natural language processing to answer image-related questions. Current models have three shortcomings: (1) Their limited interaction with other established fields hinders deep semantic analysis of both images and questions, restricting practical applications; (2) Predicted answers are often limited to single words or short phrases, lacking diversity and completeness to meet varied user needs; (3) Exclusive reliance on accuracy oversimplifies evaluation, compromising comprehensive performance assessment and model optimization. To address these challenges, we propose RCVQA, a novel reading-comprehension-based VQA model that introduces key innovations in methodology and evaluation metrics. First, dataset text paragraphs are preprocessed to remove irrelevant information, enhancing contextual focus. To ensure comprehensive evaluation, three model variants-RCVQA_P, RCVQA_C, and VQA_T-are designed and tested across four popular datasets. To address the limitations of single-answer prediction in existing methods, two innovative algorithms are developed: answer position prediction and answer content prediction, enabling multi-span answer retrieval for more comprehensive reasoning. Recognizing the limitations of accuracy as a sole evaluation metric, four novel metrics-PPR_∥, TPR_∥, H_∥-Means, and ESM_∥-are introduced to provide more comprehensive criteria for evaluating the RCVQA system. Additionally, we integrate strategies such as image captioning and local training, enhancing content understanding and ensuring data security. Experimental results demonstrate that RCVQA significantly outperforms state-of-the-art methods, achieving 1 %-7 % improvements across benchmarks: A-OKVQA (DA:+1.05 %, MC:+2.22 %), KR-VQA (+7.28 %), GQA (Val:+7.10 %, Test:+7.06 %), and OK-VQA (+4.30 %). Our implementation codes will be publicly available at: https://github.com/jianruichen/RCVQA.