Generating high-quality labels is crucial for self-supervised learning in low-light conditions, where traditional enhancement methods often struggle to balance detail enhancement and color fidelity. This paper presents a traditional image fusion approach that dynamically combines Multi-Scale Retinex (MSR) and Adaptive Histogram Equalization (AHE) outputs with the original image using an adaptive weighting strategy. The primary goal is not to compete with state-of-the-art deep learning-based enhancement methods but to produce intermediate images that can serve as effective labels for training self-supervised models without requiring ground-truth datasets. By dynamically fusing MSR and AHE outputs with the original image using adaptive brightness and color weights, the method improves structural integrity while enhancing brightness and color consistency. Experiments on standard low-light datasets demonstrate significant improvements in PSNR and SSIM compared to traditional enhancement methods. However, a visual analysis of the generated labels reveals differences in color saturation when compared to ground truth, providing insights into designing a suitable loss function for future self-supervised learning applications. It is important to note that this work does not include experiments or methods related to self-supervised learning itself; instead, it focuses on preparing high-quality labels for such approaches. Additionally, our method strikes a balance between computational efficiency and visual quality, making it suitable for real-time applications and paving the way for more robust and versatile learning frameworks.

With the rapid expansion of e-commerce and social media platforms, the prevalence of fake reviews has become increasingly problematic, misleading consumers and harming both the reputation of businesses and fair market competition. This article aims to develop a more effective technological solution to accurately identify and filter deceptive reviews, ensuring a truthful shopping and communication environment for consumers. Initially, a multifeature fusion strategy is introduced, integrating text characteristics of reviews, reviewer behavior, and product information. Through a parameterized attention mechanism, the model meticulously assigns weights to various influential features, thereby enhancing the detection of deceptive reviews. Furthermore, a composite architecture, DynDistilBERT-BiLSTM-CNN, is proposed. DynDistilBERT employs a control gate to assess the complexity of the input text and the required processing power in real-time during model forward propagation, dynamically selecting active layers within DistilBERT. This selection process is optimized with a hierarchical training strategy to minimize additional computational overhead. BiLSTM excels in processing sequential data, capturing temporal text features, while convolutional neural network focuses on identifying local text features. This approach reduces computational resource consumption for simpler tasks while maintaining high accuracy for complex tasks, recognizing both local features and contextual relationships in text. Extensive testing on the Amazon data set, compared to models like ALBERT, SpanBERT, DistilBERT, and RoBERTa, demonstrates that our model achieves accuracy improvements of approximately 5.7%, 5.2%, 5.0%, and 3.9%, with a peak accuracy of 92.6%. These findings underscore the effectiveness of the multifeature fusion strategy and the superior performance of the DynDistilBERT-BiLSTM-CNN architecture in handling complex textual data.

Image super-resolution reconstruction is one of the methods to improve resolution by learning the inherent features and attributes of images. However, the existing super-resolution models have some problems, such as missing details, distorted natural texture, blurred details and too smooth after image reconstruction. To solve the above problems, this paper proposes a Multi-scale Dual-Attention based Residual Dense Generative Adversarial Network (MARDGAN), which uses multi-branch paths to extract image features and obtain multi-scale feature information. This paper also designs the channel and spatial attention block (CSAB), which is combined with the enhanced residual dense block (ERDB) to extract multi-level depth feature information and enhance feature reuse. In addition, the multi-scale feature information extracted under the three-branch path is fused with global features, and sub-pixel convolution is used to restore the high-resolution image. The experimental results show that the objective evaluation index of MARDGAN on multiple benchmark datasets is higher than other methods, and the subjective visual effect is better. This model can effectively use the original image information to restore the super-resolution image with clearer details and stronger authenticity.

Abstract Watermark adding is one of the important means for image security and privacy protection in Internet of things (IOT) applications based on federated learning. It is often inseparable from adversarial training with watermark removal algorithms. The effect of watermark removal algorithms will directly affect the final result of watermark addition. However, the existing watermark removal algorithms have drawbacks such as incomplete image watermark removal, poor image quality after watermark removal, large demand for training data, and incorrect filling, which seriously affects the development of image information security and privacy protection in IOT applications based on federated learning. To solve the above problems, this paper proposes an improved image watermark removal convolutional network model based on deep image prior. First, we improve the U‐Net network model, using six downsamping layers and six deconvolution layers combined with deep image prior method to reduce the loss of details and perceive high‐level features, thereby improving the ability of the network to extract high‐level features of the image. In addition, we design a new type of loss function which is called stair loss, and add L1 loss and perception loss to establish new constraints. In order to verify the effectiveness of our method, a comprehensive experimental comparison was conducted on the public dataset PASCAL VOC 2012 in the same experimental environment with CGAN and the deep prior method. The experimental results show that the improved model combined with the deep image prior method can extract the high‐level feature information and can directly remove the watermark from the picture without pretraining the network, the L1 loss and perceptual loss can better retain the image structure information and speed up the watermark removal of the model, the stair loss corrects the final output more accurately by correcting the output of each layer; our method improves the learning ability of the model, and under the condition of the same training time, the image quality after watermark removal is higher, and the final watermark removal result is better, which is more suitable for distributed structure of IoT application based on federated learning.