Text Detection Method With Emphasis on Text Component Importance and Lightweight Design

Abstract

As the main carrier of information exchange, text detection is crucial in image scenes. However, the complexity of text detection tasks is high, there are many interference items, and existing text detection methods have problems such as excessive smoothing. Therefore, this study proposes an image scene text detection method based on a lightweight deep relation inference scene text detection algorithm. This algorithm simplifies the network of detection algorithms that consider the importance of text components, aiming to reduce computational complexity while addressing issues such as excessive smoothing. The experiment showed that the proposed method performed the best in accuracy, recall, and F1 value than the comparative algorithms on different datasets. For example, in the ICDAR2015 database, its accuracy, recall, and F1 values were 97.27%, 96.42%, and 96.02%, respectively. The proposed method exhibited a peak memory occupation rate of only 24.16%, with an average parameter volume of 13.57 MB over 100 tests, a value that is considerably lower than that observed in comparative algorithms. In practical applications, the proposed method consistently demonstrates optimal performance, with minimal instances of false positives or false negatives. In conclusion, the research contributes to the advancement of text detection and provides a reference for the development of lightweight networks and the optimization of performance in other computer vision tasks.

Keywords

• Text detection
• natural scenes
• excessive smoothing
• lightweight
• reduction of parameter count