Semantic segmentation is fundamental for autonomous driving, requiring dense scene understanding across diverse and continuously changing environments. Real-world deployment must overcome two major challenges: (1) dynamic environmental shifts due to weather, lighting, and geography, and (2) the inability to retain labeled source data for continual adaptation. To address these issues, we propose a novel source-free adaptive and forgetting-resilient continual unsupervised domain adaptation for semantic segmentation (FORECASS). We introduce a teacher–student based framework, using EMA (Exponential Moving Average) updating technique, to produce stable pseudo-labels during continual adaptation. Central to our method is a refiner-based error estimation model that predicts pixel-wise pseudo-label reliability during adaptation. By leveraging the error map, the model selectively focuses learning on uncertain and challenging regions, playing a critical role in mitigating catastrophic forgetting. Complementarily, a structure-aware consistency mechanism enforces semantic coherence across views, further enhancing stability during sequential adaptation. Lightweight knowledge distillation is also incorporated to smooth alignment between the pseudo-label generator and the adaptation model. We validate our framework on the challenging continual adaptation sequence GTA → Cityscapes → IDD → Mapillary, achieving state-of-the-art results over both source-dependent and source-free baselines. The code is available at: https://github.com/baqar61/FORECASS .

Human action recognition is a prominent area of research in computer vision due to its wide-ranging applications, including surveillance, human–computer interaction, and autonomous systems. Although recent 3D CNN approaches have shown promising results by capturing both spatial and temporal information, they often struggle to incorporate the environmental context in which actions occur, limiting their ability to discriminate between similar actions and accurately recognize complex scenarios. To overcome these challenges, a novel and effective approach called Context-aware 3D CNN for Action Recognition based on Semantic segmentation (CARS) is presented in this paper. The CARS approach consists of an intermediate scene recognition module that uses a semantic segmentation model to capture contextual cues from video sequences. This information is then encoded and linked to the features captured by the 3D CNN model, resulting in a comprehensive global feature map. CARS integrates a Convolutional Block Attention Module (CBAM) that utilizes channel and spatial attention mechanisms to focus on the most relevant parts of the relevant 3D CNN feature map. We also replace the traditional cross-entropy loss with a focal loss that can better deal with underrepresented and hard- to-classify human actions. Extensive experiments on well-known benchmark datasets, including HMD51 and UCF101, show that the proposed CARS approach outperforms current 3D CNN-based state-of-the-art approaches. Moreover, the context extraction module in CARS is a generic plug-and-play network that can improve the classification performance of any 3D CNN architecture. • An effective scene action recognition based on semantic contextual information. • Integration of a reliable attention mechanism to enhance feature learning. • Improving the recognition of underrepresented action classes via focal loss. • Experiments on challenging benchmarks show higher performance over SOTA techniques.

Abstract Generative Adversarial Networks (GANs) have gained prominence in computer vision, with applications that extend to image fusion. Existing fusion methods often require extensive labeled data and task-specific training, limiting their generalizability. To address these limitations, this paper presents the Reconstruction-Guided Generative Adversarial Network (R2GAN), a generic GAN-based approach designed for generic image fusion, including visible-infrared, medical, and multi-focus image fusion. The proposed R2GAN architecture consists of a primary generator to improve fusion capabilities and auxiliary generators to ensure accurate reconstruction of source image features. To optimize the model, we propose a reconstruction-guided loss function to preserve the feature distribution of the source images and improve the consistency between the fused and source images. Additionally, we introduce a semantic segmentation-guided approach to generate a comprehensive and realistic Paired Multi-Focus image dataset (PMF) to train the R2GAN model. Experimental results in multiple fusion tasks demonstrate that R2GAN delivers superior performance, outperforming state-of-the-art image fusion methods. The R2GAN framework source code is available for access on GitHub at https://github.com/CHAHI24680/R2GAN .

Semantic segmentation is a well-studied topic and one of the most challenging tasks in computer vision applications, such as autonomous driving. Deep learning approaches based on convolutional neural networks (CNN) have demonstrated exceptional success on this task in recent years. Despite this success, existing approaches are plagued by higher-order inconsistencies between the ground truth images and the ones predicted by the segmentation model. This paper proposes a novel post-processing scheme based on adversarial learning to counter these inconsistencies. Such a scheme can be combined with a variety of existing CNN-based semantic segmentation networks to improve their segmentation performances. The proposed scheme is a Two-Stream Conditional Generative Adversarial Network (TScGAN), with one stream having initial semantic segmentation masks predicted by an existing CNN, while the other stream utilizes scene images to retain high-level information under a supervised residual network structure. In addition, TScGAN incorporates a novel dynamic weighting mechanism, which leads to significant and consistent gains in segmentation performance. Several comparative tests on public benchmark driving databases, including Cityscapes, Mapillary, and Berkeley DeepDrive100K, demonstrate the effectiveness of the proposed method when used with state-of-the-art CNN-based semantic segmentation models. Furthermore, the ablation experiment proved the structural rationality of our two-stream structure. The code for TScGAN could be found at https://github.com/epan-utbm/TScGAN-for-Improving-Semantic-Predictions

Haze is one of the atmospheric image degradations that causes severe distortions to outdoor images such as low contrast, color shift, and structure damage. Due to the unique physical characteristics of haze, the quality of hazy images is not accurately assessed using general-purpose image quality assessment (IQA) approaches. Therefore, several haze-aware IQA approaches have been proposed to provide more efficient dehazing quality evaluation. These approaches extract several haze-aware features to be either combined to form a single IQA metric or fed to a regression model that predicts the dehazing quality. However, these haze-relevant features are extracted using pixel intensity , in which luminance and structure information are inseparable, leading to less correlation between such features and the type of degradation they are supposed to represent. To address this issue, we propose a singular value decomposition (SVD) based IQA metric that can effectively separate the luminance component of an image from structure. This separation offers the ability to accurately evaluate the degradation at two different levels i.e. luminance and structure. The experimental results show that our proposed SVD-based dehazing quality evaluator (SDQE) outperforms the existing state-of-the-art non-reference IQA metrics in terms of accuracy and processing time.

Generative adversarial networks (GANs) represent a significant advance in the field of deep learning for image generation problems. With their ability to generate highly realistic and diverse images, GANs are quickly becoming the preferred technique for a wide range of applications. However, GANs come with several limitations and challenges, chief among which are their collapse mode and instability during training. In this paper, we propose a novel generator GAN model that incorporates the singular value decomposition (SVD) process into the decoder module. The SVD integration allows the generator to recognize the underlying structures in the feature space, resulting in more robust and effective image generation. This is achieved by adjusting the singular values during the training process, effectively optimizing the SVD-based generator to produce images that match the ground truth. Another advantage of SVD integration is its ability to perform discriminative spatial decomposition that clearly reflects the differences between the generated features and the target features. By including SVD in the generator model, the resulting loss values are higher and less prone to gradient vanishing than conventional GANs, such as Pix2Pix. Our proposed SVD-based generator model can be integrated into any auto-encoder architecture, making it as a generic and versatile solution for various image-generation tasks. The effectiveness of the proposed SVD-based GAN in image generation has been validated in three challenging benchmarks for image restoration and visible-to-infrared image translation. Extensive experiments demonstrate the significant quantitative and qualitative improvements achieved by our SVD-based GAN compared to baseline GAN architectures. The overall system outperforms the current state of the art in all benchmarks tested.

Writer identification is experiencing a revival of activity in recent years and continues to attract great deal of attention as a challenging and important area of research in the field of forensic and authentication. In this work, we introduce a reliable off-line system for text-independent writer identification of handwritten documents. Feature engineering is an essential component of a pattern recognition system, which can enhance or decrease the classification performance. A well-designed and defined feature extraction method improves the classification task. This paper proposes, for feature extraction, an effective, yet high-quality and conceptually simple feature image descriptor referred to as Cross multi-scale Locally encoded Gradient Patterns (CLGP). The proposed CLGP feature extraction method, which is expected to better represent salient local writing structure, operates at small observation regions (i.e., connected component sub-images) of the writing sample. CLGP histogram feature vectors computed from all these observation regions in all writing samples are considered as classification inputs to identify query writers using the Nearest Neighbor Classifier (1-NN). Our system is evaluated on six standard databases (IFN/ENIT, AHTID/MW, CVL, IAM, Firemaker, and ICDAR2011) including handwritten samples in Arabic, English, French, Greek, German, and Dutch languages. Comparing the identification performance with old and recent state-of-the-art methods, the proposed system achieves the highest performance on IFN/ENIT, AHTID/MW, and ICDAR2011 databases, and demonstrates competitive performance on IAM, CVL, and Firemaker databases.

This paper presents an efficient handcrafted texture operator for texture modeling and classification. The proposed descriptor, referred to as local directional decoded ternary pattern (LDDTP), consists in encoding both directional pattern features and contrast information in a compact way based on local derivative variations. The proposed operator first computes for each pixel within its 3×3 overlapping square neighborhood, on the one hand, central edge response through the 2ndderivative of Gaussian filter, and on the other hand, eight directional edge responses using the eight Frei-Chen masks to capture more detailed information. Then, spatial relationships among the neighboring pixels through the generated edge responses are exploited independently with the help of the concepts of LTP and LDP operators to enhance the discriminative power. Finally, the produced LDDTP pattern is splitted into two distinct parts: local directional decoded ternary pattern lower ( LDDTPL) and local directional decoded ternary pattern upper ( LDDTPU), which are combined into hybrid distributions to form the final LDDTP feature descriptor. Experimental results on eight publicly available texture datasets showed that the proposed LDDTP descriptor achieves classification performance, which is competitive or better than several old and recent state-of-the-art LBP variants.