Tiny-DSOD Object Detection Using Depth-Wise Dense Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems using deep neural networks (DNNs) are computationally heavy and power-intensive, making them unsuitable for low-resource devices like smartphones, which often sacrifice speed and accuracy to reduce resource usage, resulting in a significant accuracy gap compared to full-sized counterparts.
Innovation Solution
The Tiny-DSOD framework employs a depth-wise dense block-based backbone and a depth-wise feature-pyramid network front-end, utilizing depth-wise convolutional layers and bilinear interpolation to reduce computational load while maintaining high accuracy, allowing for efficient object detection on low-resource devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks with 100+ layers and thousands of nodes are used for object detection, then detection accuracy is improved, but power consumption and computational resource usage increase significantly
Solution Approach 1:
The patent segments the deep neural network into multiple stages with progressive feature extraction. Instead of processing all features through 100+ layers, the system divides computation into hierarchical levels, extracting simple features early and complex features later, reducing overall computational load while maintaining accuracy
Solution Approach 2:
The patent applies different network depths and computational complexities to different regions of the image based on importance. High-priority regions receive deeper processing while low-priority regions use shallower processing, optimizing power consumption by allocating computational resources locally rather than uniformly across the entire image
2Measurement precision
If deep neural networks with 100+ layers are used for object detection, then detection accuracy is improved, but device complexity and resource requirements increase
Solution Approach 1:
The network is segmented into modular stages with clear functional boundaries. Each stage performs specific feature extraction tasks and outputs intermediate results that can be independently optimized, reducing overall system complexity while maintaining the benefits of deep processing
Solution Approach 2:
The patent implements dynamic computation where the network adapts its processing depth and complexity based on input characteristics. For simple images, shallower processing is used; for complex images requiring higher accuracy, deeper processing is activated, making the system dynamically adjustable rather than statically complex
3Use of energy by moving object
If reduced size neural networks are used to decrease resource usage, then power consumption is reduced, but detection speed and accuracy are sacrificed
Solution Approach 1:
The patent performs preliminary feature extraction and filtering before main detection processing. By pre-processing images to extract and filter relevant features early in the pipeline, the system reduces the computational burden on subsequent stages, enabling faster detection with reduced network size while consuming less power
Data Source
AI summary
A system, article, and method of deep supervision object detection for reducing resource usage is provided for image processing and that uses depth-wise dense blocks.