Scene Text Image Super-resolution (STISR) aims to enhance the resolution of images containing text within a scene, making the text more readable and easier to recognize. This technique has broad applications in numerous fields such as autonomous driving, document scanning, image retrieval, and so on. However, most existing STISR methods have not fully exploited the multi-scale structural and semantic information within scene text images. As a result, the restored text image quality is not sufficient, significantly impacting subsequent tasks such as text detection and recognition. Hence, this paper proposes a novel scheme that leverages multi-scale structural and semantic priors to efficiently guide text semantic restoration, ultimately yielding high-quality text images. First, a multi-scale interaction attention (MSIA) module is designed to capture location-specific details of various-scale structural features and facilitate the recovery of semantic information. Second, a multi-scale prior learning module (MSPLM) is developed. Within this module, skip connections are employed among codecs to strengthen both structural and semantic prior features, thereby enhancing the up-sampling and reconstruction capabilities. Finally, building upon the MSPLM, cascaded encoders are connected through residual connections to further enrich the multi-scale features and bolster the representational capacity of the prior. Experiments conducted on the standard TextZoom dataset demonstrate that the average recognition accuracies of three evaluators—ASTER, CRNN, and MORAN—are 64.4%, 53.5%, and 60.8%, respectively, surpassing most existing methods, including the state-of-the-art ones.

Background: Existing methods for MRI brain tumor segmentation often suffer from excessive model parameters and suboptimal performance in delineating tumor boundaries. Methods: For this issue, a lightweight MRI brain tumor segmentation method, enhanced by hierarchical feature fusion (EHFF), is proposed. This method reduces model parameters while improving segmentation performance by integrating hierarchical features. Initially, a fine-grained feature adjustment network is crafted and guided by global contextual information, leading to the establishment of an adaptive feature learning (AFL) module. This module captures the global features of MRI brain tumor images through macro perception and micro focus, adjusting spatial granularity to enhance feature details and reduce computational complexity. Subsequently, a hierarchical feature weighting (HFW) module is constructed. This module extracts multi-scale refined features through multi-level weighting, enhancing the detailed features of spatial positions and alleviating the lack of attention to local position details in macro perception. Finally, a hierarchical feature retention (HFR) module is designed as a supplementary decoder. This module retains, up-samples, and fuses feature maps from each layer, thereby achieving better detail preservation and reconstruction. Results: Experimental results on the BraTS 2021 dataset demonstrate that the proposed method surpasses existing methods. Dice similarity coefficients (DSC) for the three semantic categories ET, TC, and WT are 88.57%, 91.53%, and 93.09%, respectively.

Brain tumor detection is crucial for early diagnosis and treatment planning, as it involves automatically identifying and localizing brain tumors. However, existing methods often lack accuracy in detecting highly heterogeneous brain tumors and struggle to balance detection speed. To alleviate these issues, a novel brain tumor detection method termed channel shuffling YOLO (CS-YOLO) has been proposed, which optimizes both accuracy and detection speed. First, a depthwise separable convolution with a channel shuffling RepVGG module is designed. This module combines efficient parameter computation with robust feature extraction. It extracts deep features from images, thereby enhancing both accuracy and speed of brain tumor detection. Second, to enhance the network's performance in perceiving complex brain tumor targets, a novel convolutional multi-head self-attention module is constructed. This module learns long-range dependencies at lower resolutions, thereby improving the model's capability to recognize highly heterogeneous brain tumors. Finally, a lightweight channel shuffling convolution is designed and used to construct a lightweight residual module. This module dramatically reduces the number of parameters and the computational complexity of the model by splitting and shuffling channels, thus improving the model's learning and generalization performance. Experimental results demonstrate that the proposed method surpasses YOLOv6-L, YOLOv7, YOLOv8-L, and the latest RCS-YOLO in terms of detection accuracy and speed on the Br35H dataset. Compared to the state-of-the-art methods, the proposed CS-YOLO significantly enhances brain tumor detection accuracy and speed. Specifically, network computation in GFLOPs is reduced by 41%, FPS is increased by 14%, and accuracy AP is improved by 0.8%, achieving advanced performance.