Abstract
Direct regression-based natural scene text detection methods have already achieved promising performances. However, for the bounding box prediction, they usually utilize a fixed feature selection way to select features used to predict different components, such as the distance to the boundary or the rotation angle, which may limit selection flexibility of each component prediction and thus degrades the performance of the algorithm. To address this issue, we propose a novel method called Location-Aware Feature Selection (LAFS). It separately learns the confidence of different locations' features for each component and then selects the features with the highest confidence to form a combination of the most suitable features. In other words, LAFS uses a learnable feature selection way to flexibly pinpoint feature combinations used to predict more accurate bounding boxes. After adding LAFS, our network has a large performance improvement without efficiency loss. It achieved state-of-the-art performance with single-model and single-scale testing, outperforming all existing regression-based detectors.
Abstract (translated)
URL
https://arxiv.org/abs/2004.10999
