Text detection in Images and Videos using YOLO-NAS Small

Abstract

Text detection in natural scenes presents significant challenges due to the variability in text appearance, background clutter, and lighting conditions. Many modern text detection techniques are specifically built to recognize horizontal texts. These approaches have difficulties when dealing with real-world photos that contain text in different orientations. As a result, they are unable to meet the realistic detection needs for picture streams or films. The speed of YOLO models makes them highly useful for text identification in natural settings. The image is partitioned into a grid, with simultaneous predictions for bounding boxes and class probabilities. This method enables instantaneous detection, rendering it appropriate for dynamic settings. Their capacity to acquire hierarchical characteristics straight from data improves resilience. This paper introduces a novel approach utilizing YOLO-NAS Small, enhanced with super-gradient optimization, for efficient text detection in complex environments. The model’s architecture is designed to balance accuracy and computational proficiency, making it appropriate for real-time usage. Evaluated on a diverse dataset, our model achieved a mean Average Precision (mAP) of 0.221 at an IoU threshold of 0.50, demonstrating its capability in correctly identifying and localizing text. The model recorded a recall of 0.674, indicating a strong ability to detect most text instances.