Source-linked AI summary
Scale-Invariant Convolutional Neural Networks
Yichong Xu, Tianjun Xiao, Jiaxing Zhang, Kuiyuan Yang, Zheng Zhang
TL;DR
CNNs have limited tolerance to scale variation, and common remedies enlarge the model and use extensive scale-jittering. SiCNN uses scale-specialized columns with transformed shared filters to incorporate multi-scale processing without increasing free parameters. It learns scale-specific features and improves classification robustness, including 14.22% error on unaugmented CIFAR-10, though the study remains preliminary for larger datasets.
Problem
CNNs handle scale variation less effectively than shift variation, while common solutions enlarge the network and rely on extensive scale-jittering.
Method
SiCNN uses multi-column scale specialization, transforming shared canonical filters across columns to detect patterns at different scales without increasing free convolutional parameters.
Results
SiCNN learns features at different scales and reaches 14.22% error on unaugmented CIFAR-10, improving more than 2% absolute over the baseline CNN.
Takeaways & Limitations
The architecture provides scale robustness while complementing other optimization techniques and offering a trade-off between performance and training cost.
Takeaways & Limitations
The filter-transformation equation may have infinitely many solutions for scaling up and no exact solution for scaling down, and larger-dataset evaluation remains ongoing.
Abstract
from arXiv · showhide
Even though convolutional neural networks (CNN) has achieved near-human performance in various computer vision tasks, its ability to tolerate scale variations is limited. The popular practise is making the model bigger first, and then train it with data augmentation using extensive scale-jittering. In this paper, we propose a scaleinvariant convolutional neural network (SiCNN), a modeldesigned to incorporate multi-scale feature exaction and classification into the network structure. SiCNN uses a multi-column architecture, with each column focusing on a particular scale. Unlike previous multi-column strategies, these columns share the same set of filter parameters by a scale transformation among them. This design deals with scale variation without blowing up the model size. Experimental results show that SiCNN detects features at various scales, and the classification result exhibits strong robustness against object scale variations.
1. Introduction
CNNs perform strongly across vision tasks but handle scale variation less effectively than positional variation, motivating SiCNN’s scale-specialized, parameter-sharing architecture. SiCNN shows scale-aware behavior and strong CIFAR-10 performance without relying on a larger parameter set.
- 1. Introduction: CNNs achieve near-human performance on major vision tasks, but their scale-invariance remains weaker than their shift-invariance.This mismatch limits recognition when the same pattern appears at different input scales.
- 1. Introduction: SiCNN assigns columns to particular scales and regulates their filters through shared transformations, targeting scale robustness without substantially increasing free parameters.The design also aims to reduce dependence on extensive scale-jittering.
- 1. Introduction: Analysis finds that larger-scale columns activate for larger input patterns and that the overall system becomes less sensitive to scale variation.The reported result uses unaugmented CIFAR-10 and a simple softmax classifier.
- 1. Introduction: SiCNN’s training cost grows linearly with the number of columns, although incremental refinement can reduce this cost without significantly compromising performance.The method is presented as complementary to other performance-improvement techniques.
2. Model
SiCNN introduces scale-invariance through multiple CNN columns specialized for different scales, while tying their filters through scale transformations. The architecture propagates this relationship across layers, concatenates column outputs for classification, and handles filter transformation through linear-algebraic approximations when exact solutions are unavailable.
- 2.1. Scale-Invariance Architecture: SiCNN uses multiple convolutional columns with varying filter sizes to capture objects at unknown scales, while sharing parameters through transformed canonical filters.A canonical column retains the base filters, and scale columns transform them so the same pattern can be detected at different scales.
- 2.1. Scale-Invariance Architecture: The model concatenates top-layer feature maps from all columns and feeds the resulting feature vector to a final classifier.The simplest classification layer is a softmax layer.
- 2.2. Filters in Multiple Scales: Scale-invariance is preserved layer by layer when each filter transformation satisfies the convolution-scaling relationship, with ReLU and max-pooling maintaining the corresponding representations.Inputs at different scales produce related activations across columns, and neighboring columns can jointly respond when an object falls between their designated scales.
- 2.3. Filter Transformation: Filter transformation is formulated as a linear system using vectorized images and filters, but scaling-up yields infinitely many solutions whereas scaling-down may have no exact solution.The model uses a minimum-L2-norm solution for the infinite-solution case and approximates the original image through inverse-direction scaling for the no-exact-solution case.
- 2.3. Filter Transformation: The tied-column model is trained jointly by back-propagation, aggregating derivatives from transformed filters into the canonical filters before recomputing the column filters.The transformation from canonical filters to other scales is linear, which makes the tied-filter updates convenient.
3. Experiment Results
Experiments show that SiCNN’s transformed filters and scale-specialized columns produce scale-aware activations and more robust classification than a baseline CNN, while incremental training reduces added cost.
- 3.1. Filter Transformation for Scale-Invariance: The proposed filter transformation is more robust to image scaling than using the unchanged filter and generally outperforms direct sampling, especially when images are scaled down.The evaluation compares identity, learned transformation, and sampling-based transformation methods.
- 3.2. Multi-Column Features: SiCNN columns specialize by scale: small-filter columns respond more to small images, larger-filter columns to larger images, and successive columns peak as objects enlarge.For dogs, the 3 × 3, 5 × 5, and 7 × 7 columns reach peak responses in turn as object size increases.
- 3.3. Scale-Invariant Classification: On scaled CIFAR-10, CNN performance drops over 43.22%, whereas SiCNN drops 32.42%, indicating greater robustness to scale variation.The scaled test set combines small, middle, and large object scales created by cropping and resizing CIFAR-10 images.
- 3.4. Training results on CIFAR-10: SiCNN reduces CIFAR-10 error to 14.22% on unaugmented data, over 2% below the baseline CNN, though maxout and in-network remain better.SiCNN can also complement maxout, reducing its error from 11.68% to 11.35%.
- 3.4. Training results on CIFAR-10: A six-column CNN with independent filters overfits severely, reaching zero training error while its testing error remains around 19%.SiCNN instead ties transformed filters to canonical filters across columns.
- 3.5. Incremental Training: Incremental training nearly halves cost with comparable performance in inc-1, while inc-2 adds only 1.5% cost and still outperforms the baseline CNN.Combining incremental training with maxout reaches an 11.33% error rate.
4. Conclusion
SiCNN generalizes CNN by incorporating scale and flip invariance, improving traditional CNN results while complementing other optimization techniques. The conclusion highlights learned scale-specific features, promising performance–training-cost trade-offs, and remaining open problems for larger datasets and alternative architectures.
- SiCNN incorporates scale and flip invariance into CNN and improves traditional CNN results while complementing other optimization techniques.
- The model learns features at different scales in different columns, supporting its intended multi-scale behavior.
- The preliminary study indicates a trade-off between performance and training cost, while larger and more complex datasets remain future work.