High-quality images are of great significance for vision tasks in underwater environments. However, as light propagates through water, it is scattered and absorbed, which commonly causes issues like color distortion and loss of detail, making the capture of high-quality images challenging. To improve the quality of underwater images, we propose an underwater image enhancement method that is based on channel similarity to adaptive color correction and stationary wavelet detail enhancement. Specifically, We first innovatively introduce channel similarity values to avoid red artifacts during color correction, and finely adjust the compensation amount at the pixel level based on the intensity difference between the red and green channels. By designing a new dynamic normalization range based on channel similarity, our color correction method adaptively adjusts the dynamic range of each RGB channel’s pixel value. This accommodation for color deviations in various underwater scenes enhances the color saturation of images. Subsequently, using the stationary wavelet transform, we accurately decompose the image into low-frequency and high-frequency components. Through fine processing of the low-frequency components, we optimize detail performance and enhance the visual clarity of the underwater scene. Extensive experiments on four benchmark datasets validate that our method is state-of-the-art in underwater image enhancement, excelling in both qualitative and quantitative evaluations. Additionally, our method bolsters the precision of tasks such as keypoint matching and edge detection within the realm of image processing. The code is available at https://github.com/Zhenbo-Wang/Adaptive-Color-Correction-and-Stationary-Wavelet-Detail-Enhancement .
In underwater target detection tasks, challenges such as image blurring, complex backgrounds, and aggregation of small targets lead to problems such as difficulty in model feature extraction, target leakage, and false detection. In order to improve the accuracy, real-time performance and lightweight of underwater target detection models, we propose YOLOv7-PE: an accurate and efficient YOLOv7 improved model for underwater target detection.YOLOv7-PE is based on the single-stage target detection model YOLOv7 and separates the classification and regression tasks to be processed separately by decoupling the header design to enhance the feature extraction. We also introduce an anchor-free based design, which simplifies the target detection process, reduces the prediction time, and can adapt to targets in underwater environments. And to improve the computational efficiency, we introduce the CSPSPPF module, which reduces the computational cost of the model and improves the inference speed. In addition, we introduce the CBAM attention mechanism to enhance the feature representation in both channel and spatial dimensions. Through extensive qualitative and quantitative analyses, we verified that YOLOv7-PE has higher detection accuracy and efficient performance on the task of target detection in complex underwater environments. Relative to YOLOv7, the the average detection accuracy(mAP) of YOLOv7-PE is improved by 1.23%. Meanwhile, the Frames Per Second(FPS) of the model is improved by 1.52%, while the amount of model parameters is reduced by 6.78%. Our YOLOv7-PE model performs more accurately as well as efficiently compared to other classical target detection models.
12 November 2024 in IEEE Signal Processing Letters
In RGB-D semantic segmentation for indoor scenes, a key challenge is effectively integrating the rich color information from RGB images with the spatial distance information from depth images. However, most existing methods overlook the inherent differences in how RGB and depth images express information. Properly distinguishing the processing of RGB and depth images is essential to fully exploiting their unique and significant characteristics. To address this, we propose a novel heterogeneous dual-branch framework called HDBFormer, specifically designed to handle these modality differences. For RGB images, which contain rich detail, we employ both a basic and detail encoder to extract local and global features. For the simpler depth images, we propose LDFormer, a lightweight hierarchical encoder that efficiently extracts depth features with fewer parameters. Additionally, we introduce the Modality Information Interaction Module (MIIM), which combines transformers with large kernel convolutions to interact global and local information across modalities efficiently. Extensive experiments show that HDBFormer achieves state-of-the-art performance on the NYUDepthv2 and SUN-RGBD datasets.
Wavelet-based edge detection methods have evolved significantly over the years, contributing to advances in image processing, computer vision, and pattern recognition. This paper proposes a new local optimal spline wavelet (LOSW) and the dual wavelet of the LOSW. Then, a pair of dual filters can be obtained, which can provide distortion-free signal decomposition and reconstruction, while having stronger denoising and feature capture capabilities. The coefficients of the pair of dual filters are calculated for image edge detection. We propose a new LOSW-based edge detection algorithm (LOSW-ED), which introduces a structural uncertainty–aware modulus maxima (SUAMM) to detect highly uncertain edge samples, ensuring robustness in complex and noisy environments. Additionally, LOSW-ED unifies multi-structure morphology and modulus maxima to fully exploit the complementary properties of low-frequency (LF) and high-frequency (HF) components, enabling multi-stage differential edge refinement. The experimental results show that the proposed LOSW and LOSW-ED algorithm has better performance in noise suppression and edge structure preservation.