Improved CNN classification accuracy with the addition of shallow cascading CNNs
Vasileios Pentsos, Bijay Raj Paudel, Spyros Tragoudas, Kiriti Nagesh Gowda, Mike Schmit · 2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA) · 2021
A novel methodology of augmenting the design of an existing Convolutional Neural Network (CNN) is proposed to improve its accuracy over low accuracy classes, on any dataset. The proposed structure precedes the CNN and comprises shallow CNNs arranged in a novel cascading topology to minimize the inference time. The approach utilizes the confusion matrix of the input CNN on a specific dataset to identify sets of low accuracy classes that resemble each other with respect to the error distribution. The shallow networks operate in parallel to improve the accuracy of selected low accuracy classes, without increasing the inference time. Experimentation on benchmark datasets and established CNNs shows a significant increase, up to 36.8%, in the accuracy of selected classes over the input CNN, with practically no overhead on the inference time.