Multi-Label Cropped Image Classification via Composite CNN Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural networks (CNNs) for image classification are inefficient when processing small, cropped images due to computational demands and resource constraints on embedded systems, particularly in real-time applications like autonomous driving and driver assistance, where multiple regions of interest need classification.
Innovation Solution
Implement multi-label image classification by stacking multiple cropped images into a composite image for simultaneous processing through a CNN, modifying the flattening layer and retraining the network to generate probability vectors for each image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple cropped images are processed individually through a CNN, then classification accuracy is maintained, but computational efficiency deteriorates and processing time increases
Solution Approach 1:
The patent combines multiple cropped images into a single composite image that is processed by the CNN in one forward pass. The composite image is constructed by stacking multiple cropped images as channels or spatial regions, allowing the network to simultaneously process multiple inputs and generate classification results for all images in parallel, thereby improving throughput and reducing processing time.
Solution Approach 2:
The patent introduces a new dimension to the input data structure by organizing multiple cropped images into a composite image with extended dimensions. Instead of processing images sequentially in time, the method transforms the problem into processing a higher-dimensional composite image spatially, enabling parallel computation across multiple images within a single network pass.
2Productivity
If multiple cropped images are processed individually through a CNN, then each image receives dedicated computation, but resource consumption on embedded systems increases
Solution Approach 1:
The patent merges multiple cropped images into a single composite image input, allowing the embedded system to process multiple images through one CNN forward pass instead of multiple separate passes. This consolidation reduces the total computational workload, decreases energy consumption, and improves resource utilization on resource-constrained embedded devices.
3Productivity
If the CNN architecture is modified to handle composite images, then processing efficiency improves, but device complexity increases
Solution Approach 1:
The patent segments the composite image processing into distinct functional components: composite image construction from multiple cropped images, forward propagation through the CNN to generate class probabilities, and subsequent processing of probability vectors for each individual image. This segmentation allows the existing CNN architecture to be reused with minimal modifications, maintaining simplicity while achieving improved throughput.
Data Source
AI summary
A method for multi-label image classification in a convolutional neural network (CNN) is provided that includes forming a composite image from a plurality of clipped images, and processing the composite image by the CNN to generate a probability vector for each clipped image of the plurality of clipped images, wherein a length of a probability vector is equal to a number of classes the CNN is designed to classify.