Impacts of Layer Sizes in Deep Residual-Learning Convolutional Neural Network on Flower Image Classification with Different class sizes

This paper focuses on evaluating impacts of large -, medium - and small -size deep residual-learning convolutional neural network (DRL-CNN): ResNet50, ResNet35 and ResNet17 models on classifying Oxford-102 flowers image dataset with distinct number of 10, 50 and 102 flower classes. The Flowers image classification assessments rely on precision, recall, F1 scores and accuracy rates, averaged over 10-fold cross validation to ensure unbiased experimented results. Confusion matrix is also considered for more detail of results examination. The comparison results indicate the ResNet35 yields 0.201% and 0.706% few better recognition accuracy consecutively over ResNet50 and 17 according to 10-class dataset. For 50-class, 0.060% and 0.211% bits higher accuracy of ResNet35 than ResNet50 and ResNet17 are respectively generated. Whereas, 0.040% and 0.070% a few bits better performance of ResNet35 than ResNet50 and ResNet17 are sequentially attained on 100-class one. Decreasing rate regarding superiority of ResNet35 over ResNet50 and ResNet17 is indicated when increasing class size. However, less than 0.71% higher classification performance is indicated for ResNet35 than ResNet17 for all cases within the scope of this work. Thus, ResNet17 may be preferred to ResNet35 due to approximate 33% higher amount of parameters used in ResNet35 than ResNet17.