Pruned CNN Training Images for Small-Object Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional pruned convolutional neural networks (CNNs) struggle to learn features corresponding to small objects and textual regions in image super-resolution due to insufficient training data, leading to poor reconstruction quality and visual artifacts in high-resolution images.
Innovation Solution
A framework that extracts small objects of interest from real-world images using static detection and YOLO object detection, and overlays simulated text on cropped images to generate training data for a pruned CNN, enabling it to learn features ranging from background to textual areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional pruned CNNs are used for image super-resolution, then the network size is reduced and computational efficiency is improved, but the reconstruction quality of small objects and textual regions deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-processing input images through object detection (YOLO) and static detection to identify small objects and textual regions before super-resolution processing. These detected regions are then selectively enhanced through targeted training data generation, allowing the pruned CNN to focus computational resources on difficult-to-reconstruct areas while maintaining overall network efficiency.
Solution Approach 2:
The patent implements local quality by applying different processing strategies to different regions of the image. Specifically, small objects and textual regions are identified and processed with specialized training approaches, while other regions use standard super-resolution processing. This region-specific approach allows the pruned network to achieve high reconstruction quality where needed without requiring a fully unpruned architecture throughout.
2Loss of time
If training data is insufficient for small objects and text, then the training process is faster and requires less computational resources, but the network's ability to reconstruct these features deteriorates
Solution Approach 1:
The patent applies copying by generating synthetic training data that replicates real-world small objects and textual regions. Through text overlay generation and object detection-based data augmentation, the system creates multiple copies and variations of training samples, enabling the pruned CNN to learn robust feature representations without requiring extensive collection of real training data.
Solution Approach 2:
The patent implements parameter changes by systematically varying training data parameters such as text fonts, sizes, orientations, and object positions during training. This parameter variation allows the network to learn invariant features that generalize well across different conditions, improving reconstruction reliability without requiring a proportional increase in training data volume.
Data Source
AI summary
One embodiment provides a method comprising detecting at least one object displayed within at least one input frame of an input video. The method further comprises cropping, from the at least one input frame, at least one cropped image including the at least one object. The method further comprises generating at least one training image by overlaying simulated text on the at least one cropped image. The method further comprises providing the at least one training image to a pruned convolutional neural network (CNN). The pruned CNN learns, from the at least one training image, to reconstruct objects and textual regions during image super-resolution.