TLDR: DeepFusionNet is a novel autoencoder-based architecture designed for both low-light image enhancement and super-resolution. It achieves high image quality (SSIM 92.8%, PSNR 26.30 for low-light; PSNR 25.30, SSIM 80.7% for super-resolution) with a significantly low number of parameters (2.5 million for enhancement, 100 thousand for super-resolution), making it highly suitable for real-time applications on embedded systems.
Computer vision and image processing applications often struggle with images captured in dark or low-light conditions, especially during real-time transmission. Existing methods for brightening and coloring these images, typically using autoencoders, frequently suffer from low quality scores (SSIM and PSNR) and demand significant computational power due to their complex structures. Similarly, converting blurry, low-resolution images into sharp, high-resolution ones is a critical challenge in image processing, particularly for surveillance and security applications.
To address these persistent issues, researchers Halil Hüseyin Çalışkan and Talha Koruk from Bursa Technical University in Turkiye have developed a new architecture called DeepFusionNet. This innovative model is designed for both low-light image enhancement and super-resolution, aiming to deliver high-quality results with remarkably low computational demands.
DeepFusionNet: A Closer Look at the Architecture
DeepFusionNet is built upon an autoencoder framework, which fundamentally consists of three parts: an encoder, a bottleneck, and a decoder. The encoder compresses the input image into a compact representation, the bottleneck holds this essential information, and the decoder then reconstructs the enhanced or super-resolved image from this representation. The DeepFusionNet model draws inspiration from several well-known architectures like U-Net, ResNet18, and DenseNet, integrating their strengths in feature extraction and combination.
The encoder in DeepFusionNet is designed to extract both local and spatial information by processing inputs through CNN layers with different kernel sizes (3×3 and 5×5), a technique inspired by Inception Blocks. It utilizes the PReLU activation function, chosen for its ability to mitigate common issues like the ‘Dying ReLU’ problem and gradient loss seen in other functions, by allowing a learnable parameter for negative inputs. To further reduce the number of parameters without sacrificing performance, the architecture incorporates Depthwise Convolution (DWCNN) and GhostConv layers. An important component is the Convolutional Block Attention Module (CBAM), which helps the model focus on more important channels and spatial regions within the image, reweighting features for better processing.
The bottleneck section of DeepFusionNet efficiently reduces the number of channels using a Pointwise CNN and combines features in a manner similar to the DenseNet architecture, creating a rich latent space. The CBAM layers are also reused here, ensuring that feature importance is considered even at this compressed stage.
In the decoder, the model increases image dimensionality using Upsample layers, which are faster than traditional methods as they don’t perform convolution. It employs skip connections, a technique from U-Net, to directly transmit information from the encoder to the decoder, helping to preserve fine details and reduce gradient loss. Feature extraction in the decoder follows a pattern similar to ResNet18’s BasicBlocks, and CBAM layers are again integrated to refine features. For super-resolution tasks, a specialized layer block is added after the final CBAM layer to enlarge images by 2x.
Performance and Efficiency
DeepFusionNet’s performance was rigorously tested. For low-light image enhancement, trained on the LOL-v1 dataset, the model achieved an SSIM (Structural Similarity Index Measure) of 92.8% and a PSNR (Peak Signal-to-Noise Ratio) score of 26.30. Crucially, this was accomplished with approximately 2.5 million parameters, a significantly lower number compared to many U-Net-style architectures that can have around 30 million parameters. This low parameter count makes DeepFusionNet highly suitable for embedded systems.
For super-resolution, the DeepFusionNet-based method achieved a PSNR of 25.30 and an SSIM score of 80.7%. This super-resolution model is even more lightweight, containing only about 100 thousand parameters, which is roughly 32 times fewer than models like YOLOv8n. The low parameter count for both tasks underscores its potential for efficient real-time deployment.
When compared to other widely used autoencoder models for low-light image enhancement, DeepFusionNet demonstrated superior performance, achieving the highest PSNR of 26.30 and an SSIM of 0.92 among the compared models, including Zero-DCE++, GlowNet, EnlightenGAN, and RetinexNet.
Also Read:
- Efficient Video Upscaling: Introducing SkipSR’s Selective Processing Method
- Enhancing Visual Clarity for Smart Transportation in Challenging Weather
Conclusion
The DeepFusionNet architecture represents a significant advancement in low-light image enhancement and super-resolution. By combining an efficient autoencoder design with elements from U-Net, ResNet18, and DenseNet, and incorporating attention mechanisms, it delivers high accuracy while maintaining a remarkably low number of parameters. This efficiency makes DeepFusionNet an ideal candidate for real-time applications in various fields, including mobile photography, video surveillance, and autonomous vision systems. You can read the full research paper here: DeepFusionNet Research Paper.
