TLDR: COLI (Compressor for Large Images) is a novel framework that uses Implicit Neural Representations (INRs) to efficiently compress large images, especially in medical imaging. It addresses challenges of slow compression speed and suboptimal compression ratios in INR-based methods by introducing a multi-faceted training acceleration strategy (pretraining, mixed-precision, parallel training) and a unique post-training Hyper-Compression module. Evaluations show COLI achieves significantly higher compression ratios and faster training times while maintaining high image quality compared to existing methods.
In an era where high-resolution images are becoming increasingly common, especially in fields like medical imaging, the challenge of efficiently storing and transmitting these massive files has grown significantly. Traditional image compression methods, such as JPEG and JPEG2000, often struggle to maintain fine details at high compression rates. While newer data-driven approaches using neural networks offer improvements, they frequently face issues with generalization to new image types and scalability for extremely large images.
A promising alternative has emerged in the form of Implicit Neural Representations (INRs). Instead of storing raw pixel data, INRs learn a continuous mathematical function that maps spatial coordinates to pixel intensities. This means the image is stored as the network’s weights, offering a resolution-agnostic and potentially very compact representation. However, applying INRs to large images has its own hurdles: the compression process, which is essentially a training process, can be slow, and achieving high compression ratios without losing image quality remains difficult.
To tackle these limitations, researchers Haoran Wang, Hanyu Pei, Yang Lyu, Kai Zhang, Li Li, and Feng-Lei Fan have introduced a novel framework called COLI, which stands for Compressor for Large Images. COLI leverages Neural Representations for Videos (NeRV), an INR model originally designed for video compression, and enhances it with two key innovations: significant training acceleration and a unique post-training compression technique called Hyper-Compression.
Accelerating the Compression Process
Recognizing that INR-based compression involves training a neural network for each image, COLI implements a multi-faceted strategy to speed up this process. First, it uses a pretraining-finetuning paradigm. This means a generic NeRV model is initially trained on a large dataset of similar images. When compressing a new image, the model starts with these pre-trained weights, allowing it to converge much faster because it already understands common features. This ‘warm start’ significantly reduces the number of training cycles needed.
Second, COLI optimizes the training time per cycle. It employs mixed-precision training, which uses lower precision numbers (like FP16) for certain calculations to speed them up, while maintaining higher precision (FP32) for critical parts to ensure accuracy. Additionally, the way the network calculates its errors and performance metrics has been redesigned to be more efficient, leveraging the parallel processing capabilities of modern GPUs. This transforms what was once a sequential, time-consuming task into a more parallel and optimized one.
Third, COLI takes advantage of the high memory capacity of modern GPUs by enabling the simultaneous training of multiple NeRV networks on the same device. This parallel training significantly boosts overall throughput, especially in scenarios where many large images need to be compressed in batches, such as in hospitals.
Enhancing Compression Efficiency with Hyper-Compression
Once the NeRV model has learned to represent an image, COLI introduces a novel post-training technique called Hyper-Compression to further reduce the storage size of the network’s parameters. Since NeRV stores images as neural network weights, these weights themselves can be compressed. Unlike traditional model compression methods like pruning (removing less important connections) or quantization (reducing the precision of numbers), Hyper-Compression offers superior compression ratios without degrading the image quality or requiring additional retraining.
The core idea behind Hyper-Compression is to map high-dimensional parameter vectors (the network weights) to much more compact representations using a special mathematical function. This allows a large set of weights to be represented by a single, much smaller hyperparameter. This hyperparameter can then be further optimized and stored very efficiently, often as an integer index. This method avoids the common pitfalls of other compression techniques, such as catastrophic forgetting or the need for complex retraining procedures, making it fast and effective.
The entire COLI pipeline works by first dividing a large input image into smaller patches. Each patch is then encoded by the NeRV-based encoder, which generates neural weights for it. These weights are then passed through the Hyper-Compression module, transforming them into a compact set of hyperparameters for storage. During decompression, these hyperparameters are used to reconstruct the NeRV weights, which then generate the image patches, finally stitched together to form the complete image.
Also Read:
- Streamlining Segment Anything Models for Broader Use
- Advancing Image Generation with Vision Foundation Models as Efficient Visual Tokenizers
Real-World Performance
The effectiveness of COLI was rigorously tested on two medical imaging datasets: a cell-level electron microscopy dataset (CIL) and a CT Heart Segmentation dataset. The results demonstrate that COLI consistently achieves competitive or superior image quality (measured by PSNR and SSIM) at significantly reduced bits per pixel (bpp), which indicates higher compression efficiency. For instance, on the CIL dataset, COLI achieved a bpp of 0.17, which is notably lower than many traditional and even some learning-based methods, while maintaining good visual quality.
In terms of speed, COLI’s multi-facet acceleration strategy proved highly effective. It achieved up to 8.91 times faster training for comparable image quality levels. For example, achieving a PSNR of around 30 dB saw training time drop from 98 minutes to just 11 minutes. Even at higher quality levels (PSNR around 37-40 dB), COLI maintained a speedup of approximately 4 times. This makes INR-based image compression practical for real-world, large-scale datasets, especially in time-sensitive environments like hospitals.
COLI represents a significant step forward in large image compression, particularly for high-resolution biomedical images. By combining efficient INR encoding with a hierarchical Hyper-Compression module and a robust acceleration strategy, it offers a solution that is both computationally efficient and achieves high compression ratios without compromising image quality. This work paves the way for more practical storage and transmission of massive image data in various applications. For more technical details, you can refer to the full research paper.
