A New Fourier-Attention Guided Approach for Domain-Agnostic Text Localization

Text detection in images of adverse situations like underwater images and open day and night environments, where one can expect the effect of shaky and non-shaky cameras, is challenging. This work aims to develop a new model that can cope with the challenges of different domains, namely, underwater images, shaky and non-shaky images, and normal scene images for text detection. The approach leverages the Fourier attention and kernels to enhance feature extraction, focusing on high-frequency components associated with text edges. These features are fed to dual-stream corner detection by employing vertical and horizontal pooling for robust text detection. Additionally, we introduce a cross-star deformable convolution layer, guided by Fourier-derived information, which dynamically adapts its receptive field to achieve precise bounding box localization. Bounding box predictions are iteratively refined using heatmaps and offset adjustments. Overall, by integrating frequency-domain analysis with spatially adaptive convolutional operations, our method excels across diverse text detection scenarios without requiring domain-specific adaptations. The performance of the proposed method is demonstrated by testing on three different datasets: underwater, shaky and non-shaky images, and normal natural scene images. The results show that the proposed method achieves state-of-the-art performance compared to the existing methods.