AI Chip Feature Map Resolution via Image Cropping and Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial intelligence semiconductor solutions, particularly those using convolutional neural networks (CNNs), face challenges in maintaining image resolution due to the reduction in feature map sizes as they progress deeper in the network, making the output less useful for certain image applications.
Innovation Solution
The solution involves cropping input images into smaller subsets and merging the feature maps generated from these cropped images, using specific patterns to combine them back into larger feature maps, thereby improving image resolution and detail retention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If feature maps are processed deeper in the CNN to improve abstraction and classification, then the representational power increases, but the image resolution decreases
Solution Approach 1:
The input image is divided into multiple overlapping cropped images, each processed independently through the CNN to generate feature maps. These feature maps are then merged to reconstruct a high-resolution output feature map, thereby preserving spatial details while maintaining the benefits of deep processing.
Solution Approach 2:
The solution transitions from processing a single full-resolution image to processing multiple smaller cropped images in parallel, then combines them in a different dimensional arrangement (using merge patterns) to reconstruct the high-resolution feature map, effectively adding a spatial tiling dimension to the processing approach.
2Manufacturing precision
If the CNN model size is increased to maintain higher feature map resolution, then the image detail retention improves, but the hardware requirements and computational complexity increase
Solution Approach 1:
By segmenting the input image into smaller cropped regions, the solution enables processing of high-resolution content through the same sized CNN model, as each crop is processed independently at full resolution rather than requiring the entire high-resolution image to be processed at once.
Solution Approach 2:
The solution processes more data (multiple cropped images) than a single full-image processing approach, but each individual processing unit (crop) is smaller and requires less computational resources, allowing the use of smaller hardware while achieving equivalent or better overall resolution retention.
Data Source
AI summary
In some embodiments, a system includes an artificial intelligence (AI) chip and a processor coupled to the AI chip and configured to receive an input image, crop the input image into a plurality of cropped images, and execute the AI chip to produce a plurality of feature maps based on at least a subset of the plurality of cropped images. The system may further merge at least a subset of the plurality of feature maps to form a merged feature map, and produce an output image based on the merged feature map. The cropping and merging operations may be performed according to a same pattern. The system may also include a training network configured to train weights of the CNN model in the AI chip in a gradient descent network. Cropping and merging may be performed over the training sample images in the training work in a similar manner.