Abstract:To tackle the difficulties in fitting paired real-world data for single image deraining (SID), recent unsupervised methods have achieved notable success. However, these methods often struggle to generate high-quality, rain-free images due to a lack of attention to semantic representation and image content, resulting in ineffective separation of content from the rain layer. In this paper, we propose a novel cycle contrastive generative adversarial network for unsupervised SID, called CCLGAN. This framework combines cycle contrastive learning (CCL) and location contrastive learning (LCL). CCL improves image reconstruction and rain-layer removal by bringing similar features closer and pushing dissimilar features apart in both semantic and discriminative spaces. At the same time, LCL preserves content information by constraining mutual information at the same location across different exemplars. CCLGAN shows superior performance, as extensive experiments demonstrate the benefits of CCLGAN and the effectiveness of its components.

What problem does this paper attempt to address?

This paper attempts to solve several key problems in Single Image Deraining (SID), especially in the unsupervised learning framework. Specifically, the author points out that the main challenges faced by current unsupervised methods in generating high - quality, rain - free images are: 1. **Insufficient attention to semantic representation and image content**: Existing methods often overlook the semantic representation of image content, resulting in the inability to effectively separate the content layer and the rain layer. 2. **Problems with Cycle Consistency Loss**: Although cycle consistency loss helps maintain the reconstruction quality of the image, it also brings the problem of "channel pollution", which complicates the interaction between image components and makes it difficult to maintain content integrity and spatial accuracy during the regeneration process. 3. **Lack of effective disentanglement of semantic details**: Existing methods do not pay sufficient attention to the separation of semantic details, resulting in incomplete removal of semantic information from the rain layer, thus affecting the quality of the final image. To solve these problems, the author proposes a new Cycle Contrastive Generative Adversarial Network (CCLGAN). CCLGAN aims to improve image reconstruction quality and more effectively remove the rain layer by combining Cycle Contrastive Learning (CCL) and Location Contrastive Learning (LCL). ### Main contributions 1. **Introduction of the CCLGAN framework**: This framework can efficiently remove the rain layer and retain key content details through cycle contrastive learning and location contrastive learning. 2. **Innovative cycle contrastive learning**: It includes two cooperative branches: - **Intra - CCL**: Achieve high - quality image reconstruction in the semantic latent space while removing the rain layer in the semantic latent domain. - **Inter - CCL**: Effectively remove the rain layer in the discriminant latent space by integrating similar embeddings (rain - free information from different images) and separating dissimilar embeddings (rain - free and rain - containing information). 3. **Comprehensive evaluation**: Experimental results show that CCLGAN outperforms existing state - of - the - art methods in unsupervised deraining tasks and can be comparable to supervised or semi - supervised methods in some cases. In addition, extensive ablation studies verify the effectiveness of each component and its contribution to the overall performance. Through these improvements, CCLGAN can generate high - quality rain - free images under unsupervised conditions, solving the problems existing in existing methods.

Cycle Contrastive Adversarial Learning for Unsupervised image Deraining

Cycle contrastive adversarial learning with structural consistency for unsupervised high-quality image deraining transformer

Unpaired Deep Image Deraining Using Dual Contrastive Learning

Unsupervised Deraining: Where Contrastive Learning Meets Self-similarity

Unsupervised Deraining: Where Asymmetric Contrastive Learning Meets Self-similarity

DerainCycleGAN: Rain Attentive CycleGAN for Single Image Deraining and Rainmaking

Contrastive Unfolding Deraining Network

Deep Single Image Deraining using An Asymetric Cycle Generative and Adversarial Framework

Cycle-attention-derain: unsupervised rain removal with CycleGAN

Enhanced Attentive Generative Adversarial Network for Single-Image Deraining.

Deep single image deraining using an asymmetric cyclic generative and adversarial framework

Combining multiscale learning and attention mechanism densely connected network for single image deraining

Local and global knowledge distillation with direction-enhanced contrastive learning for single-image deraining

Removing Rain Based On A Cycle Generative Adversarial Network

Contrastive Learning Based Recursive Dynamic Multi-Scale Network for Image Deraining

SE-RRACycleGAN: Unsupervised Single-Image Deraining Using Squeeze-and-Excitation-Based Recurrent Rain-Attentive CycleGAN

Cross-domain collaborative learning for single image deraining

A Hybrid CNN-Transformer Architecture with Frequency Domain Contrastive Learning for Image Deraining

Online-updated High-order Collaborative Networks for Single Image Deraining

Detail-recovery Image Deraining via Dual Sample-augmented Contrastive Learning

Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains