Entropy-Based Post-Processing for Multi-Task DNN Misrecognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-task learning models, particularly those used in image recognition for autonomous vehicles, often misrecognize areas with lens contamination, roads, vehicles, and pedestrians under low illumination, leading to degraded performance.
Innovation Solution
A post-processing apparatus and method that generate entropy images from output images of deep neural networks within the multi-task learning model, allowing for pixel-level entropy analysis to correct misrecognitions by distinguishing between contaminated and normal areas based on entropy thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a multi-task learning model is used to simultaneously perform multiple image recognition tasks, then the model can process multiple types of information (soiling detection, road recognition, vehicle recognition, pedestrian recognition, depth estimation), but the model tends to misrecognize areas with lens contamination, roads, vehicles, and pedestrians under low illumination conditions
Solution Approach 1:
An intermediary post-processing module is introduced between the multi-task learning model and the final output. This module receives the model's output images and applies entropy-based filtering to correct misrecognitions, particularly in low-illumination conditions. The intermediary process separates reliable detections from erroneous ones without requiring the original model to be modified
Solution Approach 2:
The system incorporates feedback mechanisms where the post-processing module analyzes the output images and uses entropy values to identify and correct misrecognitions. The corrected images can then be fed back into the system for further processing or used to refine the model's performance over time
2Adaptability or versatility
If the multi-task learning model processes images with low illumination conditions, then it can operate in nighttime or dark environments, but the recognition performance degrades due to misrecognition of contaminated areas and objects
Solution Approach 1:
The system changes the parameter of image processing by introducing entropy calculation as a new dimension for analysis. Instead of relying solely on the original image pixels, the system computes entropy values for different regions and uses these entropy parameters to distinguish between valid detections and misrecognitions, particularly effective in low-illumination conditions
3Adaptability or versatility
If the model is trained to detect multiple tasks simultaneously, then it can identify soiling, roads, vehicles, pedestrians, and depth information, but the model cannot distinguish between contaminated areas and actual objects leading to misrecognition
Solution Approach 1:
The post-processing module segments the image into different regions and calculates entropy values for each segment. This segmentation approach allows the system to analyze local characteristics of different areas (soiling regions vs. object regions) and make differentiated judgments about their validity, thereby preventing loss of information about contaminated areas
Data Source
AI summary
A post-processing apparatus includes storage that stores a multi-task learning model. The post-processing apparatus also includes a controller that may generate a first entropy image from an output image of a first deep neural network (DNN) within the multi-task learning model. The controller may also generate a second entropy image from an output image of a second DNN in the multi-task learning model. The controller may post-process the output image of the first DNN based on the first entropy image and the second entropy image.