Single-Pass Object Detection via General Activation Map
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resource-constrained devices, such as smartphones and tablets, face challenges in performing generalized object detection due to the computational expense of large models required for this task, while image recognition tasks often use smaller models that are less resource-intensive.
Innovation Solution
A method utilizing a convolutional neural network (CNN) for single-pass object detection and image classification, where multi-channel data from the last convolutional layer is summed to generate a general activation map, allowing for object detection without class-specific weights, and enabling the use of this map to identify object locations within images, including interpolation and upscaling for accurate bounding box determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large models are used for generalized object detection, then detection accuracy is improved, but computational cost increases
Solution Approach 1:
The patent extracts only the essential features needed for object detection from the full image classification model. By taking out the convolutional base and removing the class-specific classification layers, the system retains the feature extraction capability while eliminating the computationally expensive classification computations, achieving object detection without full classification overhead
Solution Approach 2:
The patent uses a simplified, lightweight version of the CNN model that is computationally inexpensive compared to the full classification model. This reduced model serves as a disposable or temporary solution for object detection tasks, providing sufficient accuracy without the heavy computational burden of the complete classification architecture
2Power
If small models are used for image recognition, then computational cost is reduced, but object detection capability is limited
Solution Approach 1:
The patent makes the convolutional base universally applicable to both image recognition and object detection tasks. By training the shared convolutional layers on large-scale image classification datasets and then reusing them for object detection without task-specific fine-tuning, the system achieves multi-functionality with a single model architecture
Solution Approach 2:
The patent segments the neural network into two functional parts: a shared convolutional base for feature extraction and a simple detection head for localization. This segmentation allows the complex feature extraction to be performed once by the shared base, while the lightweight detection head provides object detection capability, achieving versatility without proportional increase in computational cost
Data Source
AI summary
In one aspect, the present disclosure relates to a method for or performing single-pass object detection and image classification. The method comprises receiving image data for an image in a system comprising a convolutional neural network (CNN), the CNN comprising a first convolutional layer, a last convolutional layer, and a fully connected layer; providing the image data to an input of the first convolutional layer; extracting multi-channel data from the output of the last convolutional layer; and summing the extracted data to generate a general activation map; and detecting a location of an object within the image by applying the general activation map to the image data.