The ever-changing underwater environment, coupled with the complex degradation modes of underwater images, poses numerous challenges to underwater image enhancement efforts. Addressing the issues of low contrast and significant color deviations in underwater images, this paper presents an underwater image enhancement approach based on color feature fusion. By leveraging the properties of light propagation underwater, the proposed model employs a multi-channel feature extraction strategy, using convolution blocks of varying sizes to extract features from the red, green, and blue channels, thus effectively learning both global and local information of underwater images. Moreover, an attention mechanism is incorporated to design a residual enhancement module, augmenting the capability of feature representation. Lastly, a dynamic feature enhancement module is designed using deformable convolutions, enabling the network to capture underwater scene information with higher precision. Experimental results on public datasets demonstrate the outstanding performance of our proposed method in underwater image enhancement. Further, object detection experiments conducted on pre- and post-enhanced images underscore the value of our method for downstream tasks.

Object detection plays a crucial role in the fields of autonomous driving, security surveillance, UAVs, etc. However, the performance of detectors can be drastically degraded by adverse weather conditions such as fog, rain, and snow. This is because detectors are usually trained on images taken in clear weather conditions but tested under adverse weather conditions. There is a domain shift problem between images captured in adverse weather and those taken in clear weather. In this paper, we propose a robust detection framework called TTSDA-YOLO (Two Training Stage Domain Adaptation YOLO), which performs well in both normal and adverse weather conditions based on YOLOv7. We design a new training strategy that fully utilizes auxiliary domains to transfer knowledge from the source domain to the target domain. This training strategy consists of two stages. In the First Training Stage, we address the disparity in feature distributions between normal weather images and adverse weather images. We use a multi-scale image-level domain adaptation module to gradually adapt the normal weather domain to the adverse weather domain. In the Second Training Stage, we make full use of the auxiliary domain by inputting it into the network as a training set. To prevent new domain shifts from being generated during the training process, we design a Backbone Regularization Module (BRM). Extensive experimental results of the proposed TTSDA-YOLO on benchmark datasets show that our approach can significantly improve the detection performance of the network in adverse weather conditions.

In recent years, with the emergence of convolutional neural networks, object detection technology has made rapid progress. However, object detection in remote sensing images remains a significant challenge due to the characteristics of remote sensing images, such as complex geographical environments, insufficient feature representation for small objects and large variation in object scale. These characteristics can lead to a significant decline in detector performance. In this paper, to address these issues, we propose an efficient detector called the Scale Sensitivity Mamba Network (SSMNet). SSMNet includes an IoU-based bounding box loss function: Area-IoU (AIoU), along with two novel Mamba-based modules: the Remote Sensing Mamba Block (RSMB) and the Sparse Feature Fusion Module (SFFM). The AIoU is used to make the detector more sensitive to the scale of the objects and to accelerate convergence on small object datasets. The RSMB leverages the powerful modeling capabilities of Mamba while maintaining the advantage of linear computational complexity to perform global context modeling on shallow features of remote sensing images. Meanwhile, the SFFM fuses the shallow features from the RSMB with the deep semantic features at the detection head to enhance the feature representation of small objects. Extensive experiments on three challenging remote sensing dataset, USOD, DIOR and DOTA, results reached 91.1% mAP50, 62.0% mAP50:95 and 24.6% mAP50:95 respectively, with only 30.4 GFLOPs, which demonstrates the effectiveness of our method. The code will be available at: https://github.com/BeStrongCode/SSMNet.

Automotive radar is one of the key sensors for intelligent driving. Radar image sequences contain abundant spatial and temporal information, enabling target classification. For existing radar spatiotemporal classifiers, multi-view radar images are usually employed to enhance the information of the target and 3D convolution is employed for spatiotemporal feature extraction. These models consume significant hardware resources and are not applicable to real-time applications. In this paper, RadarTCN, a novel lightweight network, is proposed that achieves high-accuracy online target classification using single-view radar image sequences only. In RadarTCN, 2D convolution and 3D-TCN are employed to extract spatiotemporal features sequentially. To reduce data dimensionality and computational complexity, a multi-layer max pooling down-sampling method is designed in a 2D convolution module. Meanwhile, the 3D-TCN module is improved through residual pruning and causal convolution is introduced for leveraging the performance of online target classification. The experimental results demonstrate that RadarTCN can achieve high-precision online target recognition for both range-angle and range-Doppler map sequences. Compared to the reference models on the CARRADA dataset, RadarTCN exhibits better classification performance, with fewer parameters and lower computational complexity.

Remote sensing image super-resolution (RSISR) technology can significantly improve the resolution of remote sensing images in practical applications of great significance. Remote sensing images are rich in detailed textures and multi-scale structures. Therefore, reducing the computational complexity while maintaining high-quality detail recovery becomes a key challenge. Mamba relies on a state space model of linear complexity that effectively captures long-distance dependencies but lacks recovery of detailed textures. Transformer excels at capturing global context, but has high computational overhead due to quadratic complexity. Most current methods have difficulty balancing detail recovery capability and complexity. To address these challenges, we propose a dual-branch Learnable TransMamba Hybrid Network (LTMHN) that fully integrates the strengths of state space model and attention mechanisms. Specifically, the Learnable TransMamba Block (LTMB) employs learnable adapter to dynamically fuse the attention branch with a linearly complex state space branch. This design preserves global context modeling capability while reducing computational cost. To enhance channel-wise feature representation, the Adaptive Channel Enhancement Module (ACEM) highlights important feature channels and suppresses redundant ones, substantially improving channel utilization. Experimental results show that ACEM increases the average channel activation by approximately 60% compared to a conventional Transformer. Furthermore, the Learnable Multi-Scale Overlapping Cross Attention (LMSOCA) employs a multi-scale information fusion strategy of neighboring windows to effectively enhance the recovery of detailed information such as image edges and textures. Extensive experiments on remote sensing datasets demonstrate our LTMHN could perform favorably against state-of-the-art Transformer-based and Mamba-based methods, both qualitatively and quantitatively. In brief, LTMHN achieves an average PSNR improvement of approximately 0.11 dB over the advanced method MambaIR, but only accounts for 68% and 66% of its computational cost and parameters.