Multi-Layer CNN Feature Fusion for Efficient Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection techniques in machine vision face challenges in achieving high accuracy and speed while efficiently processing images, often requiring a large number of region proposals, which can lead to increased computational burden and missed objects.
Innovation Solution
The HyperNet technology combines low-level, middle-level, and high-level convolutional feature maps from a pre-trained CNN model to extract robust Hyper Features, which are then used in a unified framework for region proposal generation and object detection, employing end-to-end joint training to share features across tasks and reduce the number of proposed regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of region proposals are generated to improve object detection accuracy, then the detection precision improves, but the computational burden increases and processing speed decreases
Solution Approach 1:
The patent segments the feature extraction process by utilizing feature maps from multiple layers of the CNN model (low-level, middle-level, and high-level features) separately. Each layer's feature map is processed independently to generate region proposals, allowing the system to leverage different levels of abstraction without requiring a single exhaustive proposal list, thus improving accuracy while managing computational load
Solution Approach 2:
The patent adds a dimensional aspect to region proposal generation by incorporating multi-layer feature maps with different spatial resolutions and semantic levels. Instead of generating proposals in a single feature space, the system operates across multiple dimensional layers of the CNN feature hierarchy, enabling more efficient proposal generation that captures objects at various scales and complexities
2Measurement precision
If a large number of region proposals are generated to improve object detection accuracy, then the detection precision improves, but the computational resources and complexity increase
Solution Approach 1:
The patent segments the computational workload by processing different CNN layers independently. Each layer's feature map is used to generate a separate set of region proposals, allowing the system to distribute computational tasks across multiple processing streams rather than handling all proposals in a single monolithic process, thereby reducing overall computational burden
Solution Approach 2:
The patent applies partial action by selectively using feature maps from specific CNN layers (low-level, middle-level, high-level) rather than processing all possible feature representations. This selective approach generates sufficient region proposals for accurate detection without the excessive computational cost of utilizing every available feature layer
3Reliability
If more region proposals are generated to improve object recall, then fewer objects are missed, but the processing time increases
Solution Approach 1:
The patent performs preliminary action by pre-processing and extracting feature maps from multiple CNN layers before region proposal generation. This advance preparation of multi-level features enables the system to quickly generate accurate region proposals without requiring extensive processing during the actual detection phase, thus improving recall while minimizing processing time
Solution Approach 2:
The patent maintains continuity of useful action by continuously leveraging the hierarchical feature representations across multiple CNN layers throughout the region proposal generation process. Instead of discrete, separate processing steps, the system continuously integrates information from different feature levels, enabling efficient and uninterrupted detection that improves recall without time loss
Data Source
AI summary
Region proposal is described for image regions that include objects of interest. Feature maps from multiple layers of a convolutional neural network model are used. In one example a digital image is received and buffered. Layers of convolution are performed on the image to generate feature maps. The feature maps are reshaped to a single size. The reshaped feature maps are grouped by sequential concatenation to form a combined feature map. Region proposals are generated using the combined feature map by scoring bounding box regions of the image. Objects are detected and classified objects in the proposed regions using the feature maps.