Weighted Joint Optimization of Convolutional UNet-Autoencoder Network for Enhanced Flower Image Classification Under Gaussian Noise

The prevalence of inter-species morphological similarities, coupled with the vast number of existing flower species poses a substantial challenge to automated flower classification. This paper proposes a Convolutional UNet-Autoencoder Network (CNNUNet), consolidated with a joint loss (JL) function, which combines Mean Squared Error (MSE) and cross-entropy with empirically derived weights, named as CNNUNet-JL. This proposed method enables multi- objective optimization for flower image classification, as exhibited here on Oxford 17 and 102 Flower image datasets. Performance's effectiveness and robustness were evaluated by various level of Gaussian noise applied to the input images. Evaluation metrics include Accuracy, F1- score, Precision and Recall are employed for classification performance measurement, comparing the proposed CNNUNet-JL, against traditional CNN and relative CNNAE-JL, using the same joint loss. CNNUNet-JL indicates outstanding superior results over those compared models in all cases. Besides, visual reconstruction performance is also assessed, along with quantitative MSE. Under Gaussian noise 0.5 and 1.0, CNNUNet-JL significantly reduced MSE loss, surpassing CNNAE- JL.