[論文レビュー] NASNet: A Neuron Attention Stage-by-Stage Net for Single Image Deraining

Qin Xu, Zhilin Wang|arXiv (Cornell University)|Dec 6, 2019

Image Enhancement Techniques参考文献 27被引用数 36

ひとこと要約

NASNetは軽量なNeuron Attentionメカニズムとステージごとのネットワークを導入し、単一画像の除雨に対して複数の雨モデルに対処する。これにより強い一般化性能と最先端の性能を達成する。

ABSTRACT

Images captured under complicated rain conditions often suffer from noticeable degradation of visibility. The rain models generally introduce diversity visibility degradation, which includes rain streak, rain drop as well as rain mist. Numerous existing single image deraining methods focus on the only one type rain model, which does not have strong generalization ability. In this paper, we propose a novel end-to-end Neuron Attention Stage-by-Stage Net (NASNet), which can solve all types of rain model tasks efficiently. For one thing, we pay more attention on the Neuron relationship and propose a lightweight Neuron Attention (NA) architectural mechanism. It can adaptively recalibrate neuron-wise feature responses by modelling interdependencies and mutual influence between neurons. Our NA architecture consists of Depthwise Conv and Pointwise Conv, which has slight computation cost and higher performance than SE block by our contrasted experiments. For another, we propose a stage-by-stage unified pattern network architecture, the stage-by-stage strategy guides the later stage by incorporating the useful information in previous stage. We concatenate and fuse stage-level information dynamically by NA module. Extensive experiments demonstrate that our proposed NASNet significantly outperforms the state-of-theart methods by a large margin in terms of both quantitative and qualitative measures on all six public large-scale datasets for three rain model tasks.

研究の動機と目的

• Address the poor generalization of deraining methods trained on single rain models
• Develop a lightweight neuron-wise attention mechanism to recalibrate feature responses
• Design a stage-by-stage unified architecture that propagates useful information across stages
• Fuse stage-level information dynamically using the Neuron Attention module
• Demonstrate superior performance across six public datasets and three rain model tasks

提案手法

• Proposes Neuron Attention (NA) mechanism using Depthwise Conv and Pointwise Conv to recalibrate neuron-wise features
• Introduces a stage-by-stage unified pattern network where later stages are guided by earlier stages
• Concatenates and fuses stage-level information dynamically through the NA module
• End-to-end training framework for deraining across multiple rain models
• Evaluates NASNet on three rain model tasks across six public datasets

実験結果

リサーチクエスチョン

• RQ1 Can a neuron-focused attention mechanism improve deraining across diverse rain models?
• RQ2 Does a stage-by-stage architecture that propagates information across stages improve generalization and performance?
• RQ3 How does the NA module compare to existing attention blocks (e.g., SE) in terms of performance and computation?
• RQ4 Is NASNet effective across multiple rain model tasks and datasets without being specialized to a single model?

主な発見

• NASNet significantly outperforms state-of-the-art methods on all six public large-scale datasets across three rain model tasks
• The Neuron Attention mechanism achieves better performance with Depthwise and Pointwise convolutions compared to SE blocks in experiments
• Stage-by-stage design enables informative guidance from earlier stages to later stages
• Dynamic concatenation and fusion of stage-level information via NA improves deraining results
• Extensive experiments demonstrate strong quantitative and qualitative improvements over competing methods