Increasing the Classification Rates of the Trained Models using Invariant Dataset Augmentation

Piotr Milczarski; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 1083-1092

In the paper, it is shown how to solve the problem of lack of rotation invariance in CNN networks using the Invariant Dataset Augmentation (IDA) method. The IDA method also allows us to increase the classification rates taking into account as an example the classification of the skin lesions using a small image set i.e. PH2 with 200 images, one large dermoscopic dataset i.e. Derm7pt with 1011 images, and one large COVID-19 dataset with 3175 images. In order to solve the problem of the lack of rotation invariance, the IDA method was used, and the data set was increased eight times or twice in an invariant way. In the research, we applied the IDA methods and compared the results of CNN networks VGG19, XN, and Inception-ResNet-v2 in three classification of features of skin lesions defined by well-known dermoscopic criteria, for example, the Three-Point Checklist of Dermoscopy or the Seven-Point Checklist for PH2 and Derm7pt. For COVID-19, CNN networks VGG19, XN, and Inceptionv3 were used. Due to IDA, classification parameters such as the true positive rate are increased by almost 20%, as well as the F1 score and Matthews correlation coefficients, opposite to type II error that has decreased significantly, i.e. the confusion matrix parameters resulted in: 98-100% precision, 98-100% true positive rate, 0.0-2.3% false positive rate, tests F1=0.95 and MCC=0.95. That general approach can provide better classification results while using already trained models.