To address the complexity of data in high-dimensional light field (LF), LF compression has emerged as a prominent research focus in recent years. However, diverse compression methods inevitably entail the loss of scene information, leading to notable distortions. To rectify this issue and recover lost information during the LF compression process, this paper introduces an innovative algorithm for enhancing the quality of compressed LF images. The proposed method addresses LF quality enhancement in two stages for any given LF viewpoint. In the initial stage, we leverage both local and global viewpoints for shallow feature extraction. Local viewpoints integrate spatial-angular contextual information through deformable convolution networks, while global viewpoints serve as auxiliary features to preserve fine details in the LF. In the subsequent stage, a dense residual network is employed for fine quality enhancement, resulting in improved LF viewpoints. Numerous experiments affirm that our approach has attained state-of-the-art performance, significantly enhancing the quality of compressed LF images.
Object classification and retrieval play important roles in many computer vision tasks. To achieve a good performance on RGB images, we propose a deep learning based multi-view method for both classification and retrieval. Different information of a static object is represented in multiple views, which makes the features extracted more stable against view variance and complex environment. To take advantage of both excellent performance of CNN architectures and rich information in multiple views, our approach is based on the Inception with Batch Normalization network for feature extraction and a post-fusion method is adopted. To reduce the effect of noise and improve computation efficiency, a denoising technology is explored to simulate infinite times corruption on the features for robust representation. The proposed method is capable of achieving high performance with a handful of training views, even better than some algorithms based on both RGB and depth modalities. The performance of the method is proved to be stable with different numbers of views as testing, and is comparable even with one testing view. Besides, the method is evaluated across datasets together with a popular Multi-view deep learning algorithms (MVCNN) and shows favourably influence can be brought by datasets transfer. This method is implemented on two kinds of benchmarks which are distinct in sizes to verify the efficiency and robustness of our method.
Current researches on Face Anti-Spoofing (FAS) predominantly focus on the detection of spoofing traces in face images. However, these methods may face limitations in generalization since spoofing traces can vary significantly across different domains. Due to poor generalization, which results in overfitting on the training data, cross-domain FAS methods always suffer dramatic degradation. In this paper, our focus is on predicting real faces rather than emphasizing spoofing traces as the primary approach to address the issue. We aim to address the issue of poor generalization by avoiding extracting spoofing traces. Based on this, a generative model named Real Face Generation Net (RFGN) is proposed to predicting real faces according to input face images, so that the final decision can be made by simply calculating the distances between original inputs and outputs. The experimental results show that our proposed method can achieve the state-of-the-art (SOTA) performance with a simple structure.
Point cloud data acquired through scanning typically exhibits sparse, non-uniform distribution, and a certain level of noise. Therefore, it is necessary to generate a dense and high-quality point cloud via up-sampling. In recent years, point cloud up-sampling techniques gain significant advantages due to the development of deep learning. In particular, most of current end-to-end up-sampling networks adopt point-wise constraints, e.g., Chamfer distance to train the up-sampling model. However, these point-wise constraints are inadequate to reduce residual noise, meanwhile would induce structural distortions. To further improve the capability of up-sampling networks, we propose a perception-wise constraints, namely PU-SSIM. Specifically, we adopt the typical full-reference point cloud quality metric to measure the structural similarity between the generated high-resolution point cloud and the ground truth. We managed to embed it into the up-sampling network, providing a plug-in capability. The experimental results indicate that the PU-SSIM can maintain the structural details, meanwhile reduce the residual noises. The proposed perception constraint is compatible to most mainstream methods, which would benefit the community to some extend.
An effective industrial inspection system enhances production efficiency, ensures product quality, reduces costs and boosts the factory’s core competitiveness throughout the entire production process. Ceramic tiles, widely utilized in building decoration, hold a significant market share. Product quality is essential not just for aesthetic appeal but also for building safety. Detecting surface defects in ceramic tiles poses formidable challenges, including complex background textures, low defect saliency and significant variations in defect scales. This presents important research opportunities in the field of computer vision. This article takes this type of difficult product as an example and starts with the ceramic tile dataset to design a template based industrial defect detection algorithm, including an image acquisition module, a template matching module and a defect detection module. A LoFTR-based template matching algorithm is proposed in this paper for the template matching module to accurately collect and match template images. Use the template images and perspective transformations to eliminate interference from background patterns after acquiring image pairs. At the same time, the system’s hardware selection scheme is established, and the YOLOv5 detection network is selected as the baseline model for the performance optimization study.
