General purpose image classification models have shown great performance on challenging benchmarks. However, classification still poses a challenge when dealing with small training sets and noisy images. Practical model usability in such scenarios requires more theoretical and empirical insights. We show that image classification accuracy on noisy data with small training sets can be improved using saliency heatmaps. We compare classification accuracies of models that use synthetic or eye-tracker generated heatmaps and models without them. We perform the comparison on synthetic and real data, where we observe substantially better accuracy with 10 to 1000 training images. Our results are consistent when using standard or pre-trained neural networks. Our findings encourage broader use of sailency maps in small data image classification with heavy noise.
|
