Surveillance Image Region Scaling for Consistent Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks face increased computation load due to low similarity in identification features of objects at different positions within detection images, especially when capturing angles vary, leading to complexity in layer design and neuron connections.
Innovation Solution
An image analysis method and device that adjusts specific areas within surveillance images to have consistent pixel dimension ratios by computing dimension ratio differences and adjusting these areas to conform to preset conditions, using an image receiver and operation processor to generate detection data for object detection networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If neural network processes objects at different positions with varying capturing angles, then identification features can be extracted, but computation load increases and identification efficiency decreases
Solution Approach 1:
The patent applies preliminary action by performing perspective transformation on detection images before feeding them to the neural network. The image processing unit transforms images from different capturing angles and positions into a unified reference perspective, pre-processing the data to reduce the complexity that would otherwise require deeper neural network layers and more computational resources.
Solution Approach 2:
The patent changes the parameter of image perspective by applying transformation algorithms that convert detection images into standardized reference views. This parameter transformation allows objects at different positions and angles to be represented uniformly, reducing the computational burden on the neural network while maintaining identification accuracy.
2Measurement precision
If more architectural layers and neuron connections are added to handle low similarity features, then identification accuracy can be improved, but device complexity and computation load increase
Solution Approach 1:
By performing perspective transformation before neural network processing, the system pre-aligns objects from different viewpoints, reducing the feature similarity challenges that would otherwise require deeper network architectures. This preliminary processing step allows for simpler, more efficient network designs while maintaining high identification accuracy.
Data Source
AI summary
An image analysis method of increasing identification efficiency is applied to an image analysis device having an image receiver and an operation processor. The image analysis method includes setting a target pixel per feet (PPF) and detecting a specific area within a surveillance image acquired by the image receiver, computing a first dimension ratio difference between the target PPF and an initial PPF of the specific area, utilizing the first dimension ratio difference to adjust the specific area so that a second dimension ratio difference between the target PPF and an adjusted PPF of the adjusted specific area conforms to a preset condition, and utilizing the adjusted specific area with the adjusted PPF to be detection data for object detection network.