CNN Object Detection via Image Concatenation for Small Target Resolution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

CNN-based object detectors face challenges in detecting small-sized objects due to reduced feature map sizes from convolutional layers, leading to inaccurate feature extraction and increased computational load when dealing with images of varying object sizes.

Innovation Solution

The method involves estimating target regions in training images, generating manipulated images by concatenating resized images, and applying convolutional and fully connected operations to integrate feature maps, reducing computational overhead by optimizing CNN operations for efficient object detection across various sizes without additional load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If convolutional layers reduce feature map size for efficient computation, then computational load is reduced, but detection accuracy of small-sized objects deteriorates

Engineering Contradiction:
Improvecomputational efficiencyVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The input image is divided into multiple sub-images (first sub-image, second sub-image, third sub-image, fourth sub-image) corresponding to different regions. Each sub-image is processed separately through the CNN, allowing small objects in different regions to be detected with sufficient feature resolution while maintaining overall computational efficiency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent processes images in multiple spatial dimensions by dividing the image into different regions (sub-images) and processing them separately. This dimensional approach allows the system to maintain high feature resolution for small objects in specific regions while avoiding the need to process the entire high-resolution image through all convolutional layers.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If object detection is performed on each cropped target region separately, then detection accuracy improves, but computational load increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

Multiple sub-images containing different target regions are concatenated together to form a single integrated input for the CNN. This merging approach allows the network to process multiple regions simultaneously in a single forward pass, avoiding the need for separate detection operations on each cropped region while maintaining detection accuracy.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The image is pre-divided into multiple sub-images corresponding to different target regions before being fed into the CNN. This preliminary segmentation allows the network to focus computational resources on regions containing objects of interest, improving detection accuracy without requiring separate processing of each region.

Inventive Principle:
Principle #10Preliminary action

3Speed

If feature map size is reduced by convolutional layers, then processing speed increases, but feature extraction accuracy for small objects deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidfeature extraction accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The input image is divided into multiple sub-images corresponding to different regions. Each sub-image is processed through the convolutional layers, allowing small objects in different regions to be detected with sufficient feature resolution while maintaining overall processing efficiency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different regions of the image (represented by different sub-images) are processed with appropriate feature extraction. This local quality approach ensures that regions containing small objects receive sufficient computational attention and feature resolution, while other regions can be processed more efficiently.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3686781B1Learning method and learning device for object detector based on CNN using image concatenation, and testing method and testing device using the same
Publication Date: 2024.11.06 STRADVISION
  • EP3686781B1 patent drawingFigure 1
  • EP3686781B1 patent drawingFigure 2
  • EP3686781B1 patent drawingFigure 3

AI summary

A method for learning parameters of an object detector with hardware optimization based on a CNN for detection at distance or military purpose using an image concatenation is provided. The CNN can be redesigned when scales of objects change as a focal length or a resolution changes depending on the KPI. The method includes steps of: (a) concatenating n manipulated images which correspond to n target regions; (b) instructing an RPN to generate first to n-th object proposals in the n manipulated images by using an integrated feature map, and instructing a pooling layer to apply pooling operations to regions, corresponding to the first to the n-th object proposals, on the integrated feature map; and (c) instructing an FC loss layer to generate first to n-th FC losses by referring to the object detection information, outputted from an FC layer.