A Graph-Based Computer Vision Pipeline Bridging Iconic Low-Level Processes and High-Level Symbolic Recognition via Agentic Retrieval-Augmented Generation

Abstract

Computer Vision and Image Understanding encompasses the full processing hierarchy from low-level iconic operations gradient computation, feature descriptor extraction, spatial frequency analysis through to the high-level symbolic processes of object recognition, scene interpretation, and semantic understanding. Despite four decades of progress, existing vision pipelines treat these two processing levels as largely independent: low-level features feed classifiers that produce high-level labels without a principled framework connecting the rich structural relationships among object classes to the recognition process itself. This paper proposes, an eight-stage Graph-Based Computer Vision pipeline(GraphCV-AgentRAG) that explicitly bridges iconic and symbolic processing through a class prototype similarity graph G = (V, E, A), providing structural relational context to a HOG-PCA-KNN recognition backbone and augmenting recognition outputs with graph-aware novelty analysis and four-agent agentic interpretation. The pipeline processes input images through: (1) image ingestion and upscaling, (2) HOG gradient orientation descriptor extraction, (3) PCA dimensionality reduction to 40-dimensional feature embeddings, (4) class prototype similarity graph construction from L2-normalised prototype cosine similarities, (5) KNN recognition (k=3, Euclidean), (6) ORB local feature matching, (7) top-k feature space retrieval, and (8) graph-aware novelty analysis with Isolation Forest and four-agent (Planner, Diagnosis, Control, Reflection) agentic output. Evaluated on the UCI Digits benchmark (10-class, N=1,797 images), GraphCV-AgentRAG achieves 100.00% accuracy, 100.00% Macro F1, and per-class OvR AUC ≥ 99.89% outperforming all eight compared state-of-the-art methods. Component ablation confirms the full pipeline contributes +21.7% accuracy improvement over raw pixel KNN, with graph structural features contributing +1.77% beyond HOG+PCA+KNN alone.