Dynamic AI Vision Platform for Context-Aware Image Recognition

Abstract

Three major challenges persist in the deployment of visual object recognition systems: (1) degradation of performance under poor imagery conditions, (2) high numbers of false positives from unfiltered, low-confidence detections, and (3) accessibility barriers preventing use by non-specialist users. In the paper presented here, we introduce our Dynamic AI Vision Platform, which addresses all three issues by combining adaptive multi-stage preprocessing, attention-augmented single-stage detection, and a confidence-calibrated RESTful web service. The preprocessing pipeline utilises Contrast-Limited Adaptive Histogram Equalization (CLAHE) for per-tile illumination normalisation, bilateral filtering for edge-preserving noise reduction, and aspect-ratio-preserving spatial resizing. In addition, a Convolutional Block Attention Module (CBAM) has been integrated into the C2f blocks of the YOLOv8n backbone to allow the network to provide a selective emphasis on discriminative spatial and channel-wise features with minimal additional parameters. A defined confidence threshold value τ = 0.40 helps avoid false positives, and an HTML5 browser native to the Flask 2.x REST API helps deliver full access to the complete pipeline. Evaluation of the MS-COCO 2017 validation split (5000 images; 80 categories) and a self-created subset of 200 adverse-condition images has resulted in a mean average precision (mAP@0.5) of 61.3%, precision of 87.4%, recall of 79.6%, and F1-score of 83.3%. Improvements are between 6.2-6.8% across all metrics compared to the baseline YOLOv8n. Average latency for API response times is 1.4 seconds under a standard CPU hardware platform; all functional test scenarios have been completed successfully. Modular architecture in the design of the platform allows for future integration of transformer-based backbones, INT8 quantization for deployment at the edge and fine-tuning specific to the domain.