High performance of optimizers in deep learning for cloth patterns detection

Abstract

In deep learning, optimization methods are an essential role. Optimizers are used to change weights and learn rates to reduce or minimize losses in a neural network. Nowadays, deep learning is widely used, especially in object detection tasks. In this case, cloth patterns are considered for object detection and assist visually impaired people. The visually impaired person had many limitations when doing their activity, not least when choosing clothes; it would be difficult without guidance or tools like Braille labels. In this research, a system was researched to detect 11 different cloth patterns (Argyle, Batik, Camouflage, Gingham, Dotted, Floral, Leopard, Solid, Striped, Zebra, and Zigzag) using RetinaNet with ResNet-152 architecture. To achieve the best model performance, compare 6 optimizers, such as Stochastic Gradient Descent, Root Mean Square Propagation, Adaptive Moment Estimation, Adaptive Delta, Adaptive Norm, and Adaptive Gradient was conducted. Each optimizer has trained with three different learning rates (1E-3, 1E-4, and 1E-5). A model with Adamax optimizer and learning rate 1E-4 was achieved with the highest accuracy with mAP (mean Average Precision) 91.28% during the training process. Based on the testing result this model was achieved precision 93.01%, recall 92.91%, F1-Score 92.79%, and accuracy 92.91%.