Image dehazing is a vital preprocessing step in computer vision, enhancing image clarity by reducing atmospheric scattering effects. However, the performance of most deep learning-based dehazing models is often limited by their reliance on fixed hyperparameters, which restricts their adaptability to diverse haze conditions. This paper proposes a novel framework that directly addresses this gap by integrating a U-Net architecture with the Rat Swarm Optimization (RSO) algorithm for dynamic and efficient hyperparameter tuning. The U-Net learns an end-to-end mapping from hazy to clear images, supported by a multi-loss strategy. Unlike traditional methods with fixed parameters, the RSO-U-Net adaptively optimizes key settings such as loss weights and learning rate using a composite fitness function based on Peak-Signal-to-Noise ratio (PSNR), Structural Similarity Index (SSIM), and Mean Squared Error (MSE). The experimental outcomes reveal that the suggested RSO-Optimized U-Net is better than traditional prior-based, CNN-based, and Transformer-based dehazing algorithms. The model is shown to achieve an average increase in performance by 43.17
YourTTS provides high-quality zero-shot voice conversion, but its deployment in latency- and memory-sensitive settings is limited by inference cost and model size. This work presents a deployment-oriented optimization of a YourTTS-based pipeline that balances inference efficiency, speaker similarity, perceptual quality, intelligibility, and prosody preservation. The proposed system integrates a modified Mel-spectrogram front-end combining standard pre-emphasis, equal-loudness weighting, and Kaiser-window Short-Time Fourier Transform, anti-aliased Snake activation in all residual blocks of the HiFi-GAN waveform generator, and Open Neural Network Exchange (ONNX)-based 8-bit integer quantization. Experiments use a fixed 56-pair English VCTK protocol and compare the original YourTTS model, a direct INT8-only variant, the full optimized pipeline, FreeVC, and Seed-VC. A component-level ablation shows that standard pre-emphasis contributes most strongly to word-error reduction, equal-loudness weighting to target-speaker similarity, and the Kaiser window to F0 preservation. The optimized model reduces RTF from 2.17 to 0.884, corresponding to a 2.45 × speedup. For 5 s of generated audio, these RTF values correspond to inference times of 10.85 s and 4.42 s, respectively, while the serialized ONNX model size decreases from 371 MB to 108 MB. Compared with direct INT8 quantization, the complete pipeline improves Target-SECS, MOS, WER, CER, V/UV Error, and F0 preservation while retaining faster-than-real-time inference. Its F0-PCC of 0.90 remains close to the original baseline value of 0.91 despite compression and acceleration. FreeVC achieves the lowest RTF, whereas Seed-VC obtains the strongest results on several quality metrics. Under the evaluated VCTK and RTX 3090 configuration, the proposed pipeline provides a practical quality–efficiency trade-off rather than universal superiority over recent voice conversion systems.
Ship detection in Synthetic Aperture Radar (SAR) images remains highly challenging due to coherent speckle noise, complex sea clutter, and significant scale variations. While Transformer-based detectors, such as RT-DETR, can achieve efficient global modeling, they often struggle to preserve weak target features under high-frequency noise conditions. To address these issues, this paper proposes the Scale-aware and Frequency-domain enhanced Denoising detection network (SFD-DETR). This framework consists of three key components. First, a Scale-Aware Enhancement (SAE) module employs 3D convolution to model cross-scale feature interactions, mitigating the loss of small target information in deep networks. Second, a Spatial-Frequency Triple Attention (SFTA) module is integrated into the feature fusion network, where an additional frequency domain branch enhances ship texture representation and mitigates speckle-induced feature interference. Finally, a Denoising Query Selection (DQS) mechanism allocates positive–negative denoising groups independently for each image under an unchanged query-slot budget, avoiding the batch-level maximum-target constraint of the original RT-DETR construction and stabilizing query matching. Experiments on the SSDD and HRSID datasets show that SFD-DETR achieves 79.3
Facial Emotion Recognition (FER) is an important component of affective computing and intelligent human-computer interaction, enabling automated understanding of human emotions through facial expressions. Despite significant progress in deep learning, video-based FER remains sensitive to the selection of representative frames, impacting both efficiency and accuracy. This paper introduces a deep learning framework designed to improve FER performance by integrating an optimized frame selection technique with a customized neural architecture. Two deep learning architectures, DeepRes-50 and AttnVision, have been used to systematically evaluate the impact of frame selection on emotion recognition. The performance of three frame selection strategies: I-Frame Extraction, Key Frame Selection, and Uniform Frame Sampling, is evaluated on four heterogeneous video emotion datasets: RAVDESS, eNTERFACE’05, PUMAVE-D, and SAMVEDNA. The experimental results show substantial variation across datasets, models, and frame-selection strategies, with accuracies ranging from 60.71
Facial Emotion Recognition (FER) has a wide range of applications in intelligent driving, clinical medicine, intelligent monitoring, virtual reality and many other fields. Despite recent progress, performance in practical applications remains challenging. Many existing models are computationally demanding, which limits their deployment in resource-constrained scenarios. In this study, we evaluate an incremental channel-attention mechanism with Local Binary Pattern (LBP) feature fusion for convolutional network models. The mechanism uses inter-layer feature variation to estimate channel importance, while LBP provides complementary texture information. Across five repeated runs, the fused model achieved 79.47 ± 0.58
