DCENet-based low-light image enhancement improved by spiking encoding and convLSTM

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light images degrade performance in various visual tasks.
• Existing low-light image enhancement (LLIE) methods often require paired data.
• Spiking-coding and convLSTM offer novel approaches to image feature extraction.

Purpose of the Study:

• To develop an unsupervised LLIE method using a simplified DCENet.
• To integrate spiking neural network coding and convLSTM for enhanced feature representation.
• To improve both subjective and objective quality of enhanced low-light images.

Main Methods:

• A simplified DCENet architecture was designed for unsupervised LLIE.
• Spiking coding methodology (intensity-to-latency) was employed.
• Convolutional LSTM (convLSTM) was utilized for feature integration.

Main Results:

• The proposed method achieved superior performance on LOL and SCIE datasets across five objective metrics (PSNR, SSIM, MSE, UQI, VIFP).
• Quantitative results exceeded the second-best method by significant margins.
• User studies confirmed the subjective visual quality improvements on non-reference datasets.

Conclusions:

• The simplified DCENet with spiking coding and convLSTM effectively enhances low-light images without supervision.
• The method demonstrates remarkable performance improvements over existing LLIE techniques.
• This approach offers a promising direction for practical low-light image enhancement applications.