Image Encoding Using Segmentation Map and Feature Vector
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression technologies, particularly those based on deep learning, face challenges in accurately representing the features of original images during image reconstruction, leading to differences in color, shape, or structure, which hinders their application in machine vision tasks that require maintaining specific image features.
Innovation Solution
The proposed method employs a segmentation map and image translation technology to generate a feature vector that includes color and texture information, allowing for efficient image encoding and decoding by separating and encoding a segmentation map image, feature vector, and residual image, and then reconstructing the original image using deep learning-based methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional image compression methods are used, then compression efficiency is improved, but image feature representation accuracy deteriorates
Solution Approach 1:
The patent divides the image into multiple semantic segments using a segmentation map that identifies different objects and regions. Each segment is then processed separately through dedicated neural network paths, allowing preservation of specific features (color, texture, shape) for each segment while achieving compression. This segmentation approach enables the system to maintain important image features for machine vision tasks while improving compression efficiency by processing only relevant features for each segment.
2Manufacturing precision
If deep learning-based image compression is used, then image reconstruction quality is improved, but machine vision task performance deteriorates
Solution Approach 1:
The patent applies different processing strategies to different regions of the image based on their semantic content. Important regions for machine vision tasks (identified through segmentation) receive enhanced feature preservation through dedicated neural network processing, while less critical regions use more aggressive compression. This local quality approach ensures that features critical for machine vision tasks are maintained with high fidelity while still achieving overall compression efficiency.
Solution Approach 2:
The patent transforms the image representation from pixel-space to feature-space using neural networks, changing the parameters from raw pixel values to semantic features (color maps, texture descriptors, shape features). This parameter transformation enables the system to compress images while preserving features that are most relevant for machine vision tasks, rather than attempting to preserve all pixel-level details that are less important for task performance.
3Measurement precision
If segmentation map and feature vector encoding is used, then feature preservation is improved, but encoding complexity increases
Solution Approach 1:
The patent performs segmentation and feature extraction as preliminary actions before the main compression process. By first identifying semantic segments and extracting their key features (color, texture, shape), the system prepares a structured representation that simplifies subsequent compression steps. This preliminary organization of image data into meaningful segments with extracted features reduces the complexity of the overall encoding process compared to attempting to compress raw pixels while preserving features.
Data Source
AI summary
Disclosed herein are a method, an apparatus and a storage medium for image encoding/decoding using a segmentation map. A feature vector for an image may be extracted using a segmentation map. The image may be encoded using the segmentation map and the feature vector. An output stream from an encoding apparatus may include a video stream and a feature stream. An input stream to a decoding apparatus may include a video stream and a feature stream. The image may be reconstructed using a reconstructed segmentation map and a reconstructed feature vector.