Cumulative Cost Pixel Classification for Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semantic segmentation methods using global information require repetitive computations to achieve high accuracy, leading to long computational times, especially when dealing with high-resolution images.
Innovation Solution
The introduction of a cumulative cost calculation using dynamic programming, which allows for the determination of object classes in each pixel through a single operation by setting scanning directions and calculating cumulative costs, reducing computational time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods are used to determine object labels using global information, then labelling accuracy is improved, but computational time increases significantly
Solution Approach 1:
The patent segments the computational process into two distinct phases: (1) a training phase that learns transition costs between object classes from training data, and (2) an inference phase that applies the learned transition costs to determine object labels. This segmentation allows the computationally intensive learning process to be performed once during training, while the actual labelling during inference requires minimal computation, thus resolving the contradiction between accuracy and computational time.
2Measurement precision
If high-resolution images are processed using conventional semantic segmentation methods, then labelling precision is improved, but the amount of computations required increases
Solution Approach 1:
The patent performs preliminary action by pre-learning transition costs between object classes during a training phase using training images. These transition costs are stored and reused during inference. When processing high-resolution images, the system directly applies the pre-learned transition costs without re-computing them, thereby maintaining high labelling precision while significantly reducing the computational complexity during actual image processing.
Data Source
AI summary
An object region identifying apparatus according to an embodiment identifies to which one of a plurality of predetermined object classes each pixel of an image belongs to label the pixel with an object type. The object region identifying apparatus includes following units. A base cost calculating unit calculates base costs of the respective object classes in each of the pixels. A transition cost estimating unit estimates a transition cost accrued when a transition between the object classes occurs between adjacent pixels in the image. A cumulative cost calculating unit calculates cumulative costs of the respective object classes in each of the pixels by accumulating the base cost and the transition cost for the respective object classes along a scanning direction set on the image. A class determining unit determines the object class of each of the pixels based on the corresponding cumulative cost.