Inception Encoder Multi-Scale Feature Extraction for Image Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image compression techniques using deep learning face challenges in achieving a balance between bit rate and distortion, particularly in extracting optimal latent representations for efficient image reconstruction.
Innovation Solution
The proposed solution involves an image compression system that utilizes a pyramidal resize module and an inception encoder network to extract features, along with a multi-scale dilated convolution unit and context/entropy models for generating probability models, which enables accurate feature extraction and reconstruction by concatenating multi-scale auxiliary information with latent representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep learning techniques are used for image compression to achieve high-quality reconstruction, then distortion is reduced, but bit rate increases
Solution Approach 1:
The patent segments the feature extraction process into multiple parallel branches (inception units with different filter sizes: 1x1, 3x3, 5x5, 7x7) that operate simultaneously on the input image. Each branch extracts features at different scales and levels of detail, which are then concatenated to form a comprehensive latent representation. This segmentation allows the system to capture diverse image characteristics efficiently, improving reconstruction quality without linearly increasing bit rate requirements.
Solution Approach 2:
The patent introduces multi-scale feature extraction by applying convolutional filters of varying sizes (1x1, 3x3, 5x5, 7x7) across different spatial dimensions. This dimensional approach allows the system to capture both fine-grained local details and broader contextual information simultaneously, creating a richer latent representation that improves compression efficiency and reconstruction quality without proportionally increasing bit rate.
2Measurement precision
If multiple feature extraction operations are performed to obtain competitive latent representations, then reconstruction accuracy improves, but computational complexity increases
Solution Approach 1:
The patent merges multiple feature extraction operations into a unified inception unit structure where parallel convolutional branches process the input simultaneously. The outputs from different filter sizes (1x1, 3x3, 5x5, 7x7) are concatenated and processed through a shared subsequent network. This merging approach achieves comprehensive feature extraction while reducing redundant computations compared to sequential processing, thereby improving reconstruction accuracy without linearly increasing computational complexity.
Data Source
AI summary
Embodiments of this disclosure provide an image coding apparatus, a probability model generating apparatus and an image decoding apparatus. A processor is to perform feature extraction on an input image to obtain first feature maps of N channels; to perform feature extraction on the input image with a size of the input image being adjusted K times, to respectively obtain second feature maps of N channels; and to concatenate the first feature maps of the K×N channels with the second feature maps of K×N channels to output a concatenated feature maps of channels. Hence, features of images may be accurately extracted and more competitive latent representations may be obtained.