Object Detection With Dilated Convolution for Multi-Scale Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolutional neural networks are cumbersome and inefficient, particularly in tasks requiring accurate detection of objects with varying scales and detailed localization.
Innovation Solution
A method combining dilated convolution layers with pooling layers to enhance the receptive field while maintaining detailed local information, followed by up-sampling and concatenation to improve object detection accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional CNN pooling operations are used for object detection, then computational efficiency is improved, but gridding artifacts and information loss occur
Solution Approach 1:
The patent introduces dilated convolution layers as intermediary components between pooling layers. These dilated convolution layers with expansion factors (e.g., 2, 4, 8) act as mediators that preserve local information and reduce gridding artifacts while maintaining computational efficiency. The dilated convolutions fill in the gaps created by pooling operations, providing a bridge between the downsampling operation and the final detection output.
Solution Approach 2:
The patent changes the parameters of the convolution operation by introducing dilation rates (expansion factors). Instead of using standard convolutions with fixed kernel sizes, the system uses dilated convolutions where the effective receptive field is expanded by inserting gaps between kernel elements. This parameter change allows the network to maintain higher resolution features while still achieving computational efficiency through pooling operations.
2Speed
If multiple pooling operations are applied to increase receptive field, then processing speed is improved, but local information is lost
Solution Approach 1:
The patent segments the feature extraction process into distinct stages: pooling operations for downsampling and dilated convolution operations for local feature preservation. Each dilated convolution layer is configured with a specific expansion factor that segments the receptive field into different scales, allowing the network to process information at multiple resolutions simultaneously without losing local details.
Solution Approach 2:
The patent adds the dimension of dilation rate to the convolution operation. By introducing expansion factors (1, 2, 4, 8), the system creates a multi-dimensional feature space where local information is preserved across different scales. This dimensional expansion allows the network to maintain detailed local features while still achieving the receptive field coverage needed for fast processing.
3Device complexity
If standard convolution layers are used, then device complexity is reduced, but detection precision for small objects deteriorates
Solution Approach 1:
The patent makes the dilated convolution layers universal by configuring them with multiple expansion factors that can handle different object scales. The same dilated convolution module can detect both small objects (with lower expansion factors) and large objects (with higher expansion factors), eliminating the need for separate detection heads for different object sizes and maintaining relatively simple device complexity.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A computer implemented method for object detection the following steps carried out by computer hardware components: determining an output of a first pooling layer based on input data; determining an output of a dilated convolution layer, provided directly after the first pooling layer, based on the output of the first pooling layer; determining an output of a second pooling layer, provided directly after the dilated convolution layer, based on the output of the dilated convolution layer; and carrying out the object detection based on at least the output of the dilated convolution layer or the output of the second pooling layer.