Natural Scene Text Detection Algorithm via Internal Feature Enhancement and Adaptive Cross Fusion

In recent years, convolutional neural networks (CNNs) widely used in text detection tasks can effectively predict high-frequency details. However, most CNNs exhibit limited capability in jointly processing low-frequency semantic information and high-frequency localization information, which may reduce the naturalness and fidelity of the reconstructed text to some extent, thereby affecting the performance of downstream tasks. To address these issues, we propose a natural scene text detection algorithm via internal feature enhancement and adaptive cross fusion (IFEACF). Specifically, by splitting and classifying feature maps in the channel dimension, the Internal Feature Enhancement Residual Network (IFE-ResNet) is constructed to improve the internal correlation of features and enable the network to focus on information-rich regions. In feature fusion, the adaptive cross fusion (ACF) module is proposed to fully integrate the detailed information of shallow features with the semantic information of deep features, enhancing the fusion effect of features from non-adjacent layers. Furthermore, the spatial perception fusion (SPF) module is designed to adaptively fuse feature maps of different scales, which generates accurate RoI feature to improve the text detection performance of the model. The superiority of the presented method has been demonstrated through experiments conducted on the ICDAR2015 dataset, showing that precision, recall, and F-measure surpass the baseline PSENet by 4.01%, 0.91%, and 2.34% respectively. Moreover, both precision and F-measure of the proposed algorithm outperform those of the state-of-the-art methods.