With advanced neural network techniques, language models can generate content that looks genuinely created by humans. Such advanced progress benefits society in numerous ways. However, it may also bring us threats that we have not seen before. A neural text detector is a classification model that separates machine-generated text from human-written ones. Unfortunately, a pretrained neural text detector may be vulnerable to adversarial attack, aiming to fool the detector into making wrong classification decisions. Through this work, we propose $\mu$Attacking, a mutation-based general framework that can be used to evaluate the robustness of neural text detectors systematically. Our experiments demonstrate that $\mu$Attacking identifies the detector’s flaws effectively. Inspired by the insightful information revealed by $\mu$Attacking, we also propose an RR-training strategy, a straightforward but effective method to improve the robustness of neural text detectors through finetuning. Compared with the normal finetuning method, our experiments demonstrated that RR-training effectively increased the model robustness by up to $11.33\%$ without increasing much effort when finetuning a neural text detector. We believe the $\mu$Attacking and RR-training are useful tools for developing and evaluating neural language models.

Pneumonia is a severe inflammation of the lung that could cause serious complications. Chest X-rays (CXRs) are commonly used to make a diagnosis of pneumonia. In this paper, we propose a deep-learning-based method with spatial attention superposition (SAS) and multilayer feature fusion (MFF) to facilitate pneumonia diagnosis based on CXRs. Specifically, an SAS module, which takes advantage of the channel and spatial attention mechanisms, was designed to identify intrinsic imaging features of pneumonia-related lesions and their locations, and an MFF module was designed to harmonize disparate features from different channels and emphasize important information. These two modules were concatenated to extract critical image features serving as the basis for pneumonia diagnosis. We further embedded the proposed modules into a baseline neural network and developed a model called SAS-MFF-YOLO to diagnose pneumonia. To validate the effectiveness of our model, extensive experiments were conducted on two CXR datasets provided by the Radiological Society of North America (RSNA) and the AI Research Institute. SAS-MFF-YOLO achieved a precision of 88.1%, a recall of 98.2% for pneumonia classification and an AP${}_{50}$ of 99% for lesion detection on the AI Research Institute dataset. The visualization of intermediate feature maps showed that our method could facilitate uncovering pneumonia-related lesions in CXRs. Our results demonstrated that our approach could be used to enhance the performance of the overall pneumonia detection on CXR imaging.

Image blur, often caused by camera shake and object movement, poses a significant challenge in computer vision. Image deblurring strives to restore clarity to these images. Traditional single-stage methods, while effective in detail enhancement, often neglect global context in favor of local information. Yet, both aspects are crucial, especially in real-life scenarios where images are typically large and subject to various blurs. Addressing this, we introduce CNB Net, an innovative deblurring network adept at integrating global and local insights for enhanced image restoration. The network operates in two stages, utilizing our specially designed Convolution and Normalization-Based Block (CNB Block) and Convolution and Normalization-Based Plus Block (CNBP Block) for multi-scale information extraction. A progressive learning approach is adopted with a Feature Active Selection (FAS) module at the end of each stage that captures spatial detail information under the guidance of real images. The Two-Stage Feature Fusion (TSFF) module reduces information loss caused by downsampling operations while enriching features across stages for increased robustness. We conduct experiments on the GoPro dataset and the HIDE dataset. On the GoPro dataset, our Peak Signal-to-Noise Ratio (PSNR) result is 32.21 and the Structural Similarity (SSIM) result is 0.950; and on the HIDE dataset, our PSNR result is 30.38 and the SSIM result is 0.932. Our results exceed other similar algorithms. By comparing the generated feature maps, we find that our model takes into account both global and local information well.

Deep learning-based image deblurring techniques have made great advancements, improving both processing speed and deblurring efficacy. However, existing methods still face challenges when dealing with complex blur types and the semantic understanding of images. The segment anything model (SAM), a versatile deep learning model that accurately and efficiently segments objects in images, facilitates various tasks in computer vision. This article leverages SAM’s proficiency in capturing object edges and enhancing image content comprehension to improve image deblurring. We introduce the edge-sensitive focusing encoder (EFE) module, which utilizes masks generated by the SAM framework and re-weights the masked portion following SAM segmentation by detecting its features and high-frequency information. The EFE module uses the masks to locate the position of the blur in an image while identifying the intensity of the blur, allowing the model to focus more accurately on specific features. Masks with greater high-frequency information are assigned higher weights, prompting the network to prioritize them during processing. Based on the EFE module, we develop a deblurring network called the edge-sensitive focusing encoder-based convolution–normalization and attention network (EFE-CNA Net), which utilizes the EFE module to enhance the deblurring process, employs an image-mask decoder to merge features from both the image and the mask from the EFE module, and incorporates the CNA Net as its base network. This design enables the model to focus on distinct features at various locations, enhancing its learning process through the guidance provided by the EFE module and the blurred images. Testing results on the RealBlur and REDS datasets demonstrate the effectiveness of the EFE-CNA Net, achieving peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics of 28.77, 0.902 (RealBlur-J), 36.40, 0.956 (RealBlur-R), 31.45, and 0.919 (REDS).

Most existing super-resolution (SR) methods are usually achieved through fully supervised means with massive training samples, and assume the degradation of low-resolution images corresponding to high-resolution images is fixed. However, the degradation process of real-world images is often more complex. Therefore, these models often perform poorly when dealing with low-resolution (LR) images with unknown degradation. In addition, runtime is also an important factor in deploying image super-resolution models, especially on devices with limited resources. However, the runtime of existing zero-shot blind super-resolution models is not ideal. In this paper, we propose an efficient zero-shot network (EZSN) for blind super-resolution. Specifically, we propose a high-frequency feature extraction block (HFFEB), which speeds up network inference by stacking highly optimized convolution and activation layers and reducing the use of feature fusion. In addition, we also propose an enhanced residual block (ERB) for extracting more features to improve model performance. The experimental results indicate that our proposed EZSN method has a significant advantage in terms of runtime compared to previous ZSSR methods when dealing with benchmark datasets that have unknown degradation.