CNN Object Detection Device for Small Target Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic optical tracking systems using deep learning struggle to detect small objects in real time due to limitations in the internal layer arrangement of convolutional neural networks (CNNs), leading to increased data processing and potential failure in detecting small objects.
Innovation Solution
The method involves dividing images into local areas based on the size of the CNN's convolutional layer, selecting either the moving saliency map (MM) or maximum bounding box (MB) mode, and inputting these areas to multiple CNN channels to detect small objects efficiently, with the MM mode using saliency maps and the MB mode using bounding box maps to set object detection targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the internal layer arrangement of the CNN is extended to detect small objects, then the detection capability for small objects is improved, but the amount of data to be processed becomes vast and real-time detection cannot be achieved
Solution Approach 1:
The patent divides the input image into multiple local areas (e.g., 4 quadrants) and processes each local area separately through the CNN. This segmentation allows the system to maintain a fixed, efficient CNN architecture while improving small object detection capability by focusing computational resources on specific regions rather than processing the entire image at high resolution through deep layers.
Solution Approach 2:
The patent introduces a new processing dimension by combining local area extraction with multi-scale feature extraction. Instead of extending the CNN depth to detect small objects, the system creates a new dimension of operation by processing multiple local areas at different scales simultaneously, achieving both small object detection and real-time performance.
2Measurement precision
If the image resolution is increased to SXGA or higher, then the detection resolution is improved, but the deep learning based system still cannot detect small objects of predetermined size or less
Solution Approach 1:
The system segments the high-resolution image into multiple local areas and processes each through the CNN. This allows the high-resolution image data to be maintained while the CNN processes manageable portions, enabling small object detection that would be impossible when processing the entire high-resolution image through deep CNN layers.
Solution Approach 2:
The patent applies local quality by processing different local areas of the image with appropriate computational resources. Each local area receives focused processing to maintain high detection quality for small objects within that region, while the overall system maintains efficiency by not processing the entire high-resolution image uniformly.
3Adaptability or versatility
If the CNN processes the entire image to detect all objects, then the comprehensive detection coverage is improved, but the processing time increases and real-time detection is lost
Solution Approach 1:
The patent segments the image into multiple local areas that can be processed in parallel or sequence. This segmentation maintains comprehensive detection coverage by ensuring all regions are processed while significantly reducing the time required compared to processing the entire image as a single unit through the CNN.
Solution Approach 2:
The system applies partial action by processing only the necessary local areas through the CNN rather than the entire image. This selective processing maintains adequate detection coverage for practical purposes while dramatically reducing processing time to achieve real-time performance.
Data Source
AI summary
Disclosed are an object detection device and a control method. A method for controlling an object detection device comprises the steps of: receiving one image; dividing the received image into a predetermined number of local areas on the basis of the size of a convolutional layer of a convolution neural network (CNN); identifying small objects at the same time by inputting a number of the divided local areas corresponding to the number of CNN channels to each of a plurality of CNN channels; sequentially repeating the identifying of the small objects for each of the remaining divided local areas; selecting MM mode or MB mode; setting an object detection target area corresponding to the number of CNN channels on the basis of the selected mode; and detecting the small objects at the same time by inputting each set object detection target area to each of the plurality of CNN channels.