Image deraining typically involves synthesizing low-quality degraded data for training using a predefined degraded model of a single weather condition. While in real world scenarios, varying rain intensities result in different sizes and densities of raindrops and rain streaks, increasing the complexity of image degradation. In this paper, we proposed a multi-stage deraining framework based on pre-trained diffusion model, it can efficiently perform the rain removal task under a variety of weather situations. We diffuse degraded images into a noisy state where various types of degradation are transformed into Gaussian noise. Then, during the denoising process, the low-frequency information of the image is replaced through iterative refinement, guiding the pre-trained diffusion model for image reconstruction. Our method effectively utilizes the generative priors in diffusion models and avoid the computational burden of retraining conditional diffusion models. Experimental results on four rainy degradation image datasets show its robustness to different types and severities of degradation (such as raindrops and rain streaks). Compared to recent deraining algorithms, our method achieves a maximum improvement of 0.96 dB (3.5%) in PSNR and 0.021 dB (2.7%) in SSIM for the restored images.

Macular disease has become one of the main causes of blindness. The application of optical coherence tomography (OCT) assisting ophthalmologists to analyze them is essential in clinical diagnosis and treatment. Recently, the two-stage attention-based method was proposed to classify and segment the lesion area. However, the knowledge learned by the classification task cannot be transferred to the segmentation task in the two-stage method. In this paper, we propose a novel one-stage attention-based method for retinal OCT image classification and segmentation, simultaneously the two tasks can promote each other. Specifically, attention map obtained by Convolutional neural networks (CNN) classifiers are applied for guiding the network to achieve more accurate classification. For gaining stronger discrimination ability, we apply the priori knowledge of grayscale difference to generate pseudo ground truth for constraining the attention map. Experimental results on the UCSD dataset demonstrate the efficiency of the proposed network.

Image aesthetics quality evaluation, which allows computers to judge "beauty and ugliness", is widely used in fields of image recommendation and image editing etc.. Most aesthetic evaluation methods can only output one type of evaluation result, thus, the scope of their application scenarios is limited. To solve this problem, we proposed a multi-task based aesthetic evaluation system, which can output the image style label and three forms of the aesthetic evaluation results for a image. The training procedure is divided into two stages to realize the aesthetic evaluation tasks from coarse to fine. Experiments on AVA dataset show that it can accomplish an efficient and comprehensive image aesthetic quality evaluation.

In this paper, we propose an interactive image retrieval method based on interactive image segmentation and relevance feedback. For testing the performance of the algorithm, we built an image database by web crawlers, and added a background label to each image by histogram analysis. For image retrieval, an interactive image segmentation scheme based on GrabCut has been applied to get the region of interest (ROI), and then we use an automatic labeling method to get the training samples of relevance feedback, and then incorporate the background labels into the similarity measurement to decrease the influence of clutters. The experimental results show that this method can reduce the influence of image background on image retrieval, and optimize the search results by the feedback of users.

Human head pose estimation is a hot topic in computer vision field, which can be used in video surveillance, Human Computer Interaction and so on. Active Shape Model is a template matching method, which is suitable for object localization and point based feature extraction. In this paper, we propose an algorithm based on Active Shape Model for head pose estimation. In the proposed algorithm, we firstly use Active Shape Model to estimate 2D face feature points of the target human head, then we adopt Relevant Vector Machine to evaluate head pose based on the extracted feature points. Experiments on CAS-PEAL-R1 dataset show that the proposed algorithm has great potential in estimating head pose with small yaw angle.

In this paper, a new method for the estimation of 3D human body poses from monocular images is proposed. Histograms of oriented gradients are used as the features for modeling human body poses. Human body poses are represented as 3D limb angles, which can remove the structure information from pose vector. Relevance Vector Machine is used to infer the mapping from image features to body poses. Experiments show that the proposed method is robust to camera views and can lead more accurate results than other pose estimation methods.

Particle filter is a popular stochastic tracker for object tracking. In articulated human body pose tracking, lots of work focuses on increasing sampling efficiency by incorporating optimization algorithm into particle filter. In this study, we propose a modified optimization based particle filter algorithm for pose tracking. The new algorithm can maintain the diversity of particle set by using a suppression scheme. Experimental results show that the proposed method can cope with multi-modality and can obtain more accurate estimation than other optimization based particle filter methods.