A Scene Tibetan Text Detection by Combining Multi-scale and Dual-Channel Features

Tibetan text detection in scenes plays a vital role in various applications, including image search, real-time translation, and the preservation of Tibetan cultural heritage. However, recognizing Tibetan text in natural scene images is a challenging task due to factors such as variable fonts, complex backgrounds, and poor imaging conditions. In this study, we present a novel approach called Multi-Scale Dual-Channel Feature Fusion (MDFF) for Tibetan scene text detection. Our method aims to accurately infer text in complex scenes by leveraging multi-scale interactions between texts. MDFF incorporates a feature pyramid network with skip connections, enabling the fusion of features at different scales in a hierarchical manner. Additionally, we employ a dual-channel attention (DCA) mechanism to capture rich interactions between text instances while mitigating the impact of background noise. Experimental results on the scene Tibetan text detection database (STTDD) demonstrate the effectiveness of MDFF, achieving an impressive F1 score of 85.20%. Our proposed method outperforms the baseline model by 5 percentage points and surpasses the performance of six state-of-the-art methods in single Tibetan text detection.