Neural Network Object Detection Using Multi-Channel Contour Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting multiple target objects in an image using neural networks struggle to distinguish between adjacent objects, leading to recognition as a single target object, and increasing training data does not effectively solve this issue.
Innovation Solution
A method involving a neural network-based detection model that utilizes multi-channel Ground Truth data, including contour, adjacent area, and target area information, to generate single-channel output data by adjusting values in the detection model, and incorporates polarized images and synthesized images for improved object separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a neural network-based detection model is used to detect target objects, then the automation and speed of object detection are improved, but when multiple target objects are located adjacent to each other, they cannot be distinguished and are recognized as one target object
Solution Approach 1:
The patent applies segmentation by dividing the detection task into multiple specialized networks: a first detection network for detecting target objects and a second detection network for detecting contour lines. This segmentation allows each network to specialize in specific aspects of detection, improving the ability to distinguish adjacent objects while maintaining automation.
Solution Approach 2:
The patent introduces a new dimension by adding contour line detection as a separate detection layer. Instead of relying solely on object detection, the system adds contour line detection results as an additional dimension of information, enabling better separation and identification of adjacent target objects through the combined analysis of both detection dimensions.
2Measurement precision
If the number of training data is increased to improve detection accuracy, then more target objects can be detected, but it cannot solve the problem of distinguishing adjacent target objects and requires a separate learning method
Solution Approach 1:
The patent segments the detection system into two specialized networks with different functions. The first network focuses on target object detection while the second network focuses on contour line detection. This segmentation allows each network to be optimized for its specific task, achieving high accuracy for distinguishing adjacent objects without requiring a single complex network.
Solution Approach 2:
The patent creates a multi-functional detection system where the first detection network handles target object identification and the second detection network handles contour line detection. By combining the results of these two networks, the system achieves universal detection capability that can handle both general object detection and the specific challenge of distinguishing adjacent objects.
3Device complexity
If a single detection network is used, then the device complexity is reduced, but it cannot separately detect and distinguish multiple adjacent target objects
Solution Approach 1:
The patent divides the detection functionality into two separate networks, each with specialized responsibilities. This segmentation increases functional capability while keeping each individual network relatively simple, balancing device complexity with detection precision for adjacent objects.
Solution Approach 2:
The patent introduces contour line detection results as an intermediary element that mediates between the target object detection and the final identification of adjacent objects. The contour lines serve as intermediate information that helps distinguish boundaries between adjacent target objects, enabling the system to maintain simplicity while achieving high precision.
Data Source
AI summary
A method of detecting a target object performed by a computing device including at least one processor according to an exemplary embodiment of the present disclosure may include: receiving an input image; and generating first result information related to an area corresponding to a target object from the input image based on a trained neural network-based detection model.