CNN Small Object Detection via Two-Stage Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural networks are inefficient in detecting small objects, such as distant pedestrians, due to high miss rates and slow processing times, making them impractical for self-driving cars, where timely detection is critical.
Innovation Solution
An enhanced deep neural network approach involving a two-stage training process for convolutional neural networks (CNNs), where a first CNN is trained on small objects and re-trained using images of larger objects, minimizing the difference in feature extraction between the two networks to improve detection accuracy of small objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep neural networks are used for object detection, then detection capability is provided, but processing speed is slow (seconds or minutes per image)
Solution Approach 1:
The training process is segmented into two distinct stages: first training a CNN on small objects, then re-training with large objects while minimizing feature extraction differences. This segmentation allows the network to specialize in small object detection while incorporating general object features, achieving both high detection capability and faster processing speed for real-time applications.
Solution Approach 2:
The patent changes the training parameters by introducing a two-stage training approach with different object size inputs. The first stage trains on small objects (30-50 pixels) to specialize in detecting distant pedestrians, while the second stage re-trains on large objects to incorporate general object features. This parameter change in training methodology resolves the contradiction between detection capability and processing speed.
2Productivity
If conventional methods are used for detecting small pedestrians, then processing is performed, but miss rate is high (approximately 80% for 30-50 pixel objects)
Solution Approach 1:
The patent performs preliminary action by first training the CNN on small objects before re-training on large objects. This preliminary training stage enables the network to learn specific features of small, distant pedestrians, significantly reducing the miss rate from 80% to much lower levels while maintaining practical processing speed for self-driving applications.
Solution Approach 2:
The patent implements feedback by minimizing the difference in feature extraction between the first CNN (trained on small objects) and the second CNN (trained on large objects). This feedback mechanism ensures that the re-trained network maintains its ability to detect small objects while incorporating general object detection capabilities, thereby reducing miss rates without sacrificing detection coverage.
3Measurement precision
If features of small pedestrians are analyzed, then detection is attempted, but features appear blurry making discrimination difficult
Solution Approach 1:
The patent performs preliminary training on small objects to enhance feature extraction capabilities for blurry, distant pedestrians. This preliminary action allows the network to learn robust features from low-quality inputs, improving measurement precision for feature discrimination despite the inherently blurry nature of small object features in captured images.
Solution Approach 2:
The patent changes the training parameters by using zoomed-in images of small objects during the first training stage. This parameter change allows the network to learn from enhanced, higher-resolution features during training, improving its ability to discriminate blurry features in actual small object detection scenarios while maintaining reliability.
Data Source
AI summary
Various methods are provided for training and subsequently utilizing a convolutional neural network (CNN) to detect small pedestrians (e.g., pedestrians located away a large distance). One example method may comprise performing a first training stage in which a first CNN is trained to detect objects of a first size, the first CNN trained using a first set of images comprised of objects of the first size, and configured to output a first set of parameters, performing a second training stage in which a second CNN is trained using a second set of images, the second set of images comprising objects of a second size, and the first CNN is initialized with the first set of parameters and is re-trained using the second set of images, and determining parameters of the first CNN by minimizing error between the first CNN and the second CNN.