Context-Attention Architecture for Multi-Oriented Text Detection

The detection of scene text constitutes a formidable challenge within the realm of computer vision. Serving as the preliminary step before text recognition, it holds diverse applications in everyday tasks such as translation, self-driving, and more. This article introduces a novel pipeline for text detection that leverages contextual information within images to enhance detection accuracy and reduce false positives. Our architecture is rooted in semantic segmentation and draws inspiration from the Pyramid Scene Parsing Network. To further enhance its capabilities, we have incorporated a channel-space attention submodule. Additionally, we propose a new loss, inspired by the dice loss function, designed to place a stronger emphasis on mitigating false detections and thereby improving overall detection performance. Our model has been tested on various datasets and has successfully achieved state-of-the-art performances.