An Efficient Grocery Detection System Using HYOLO-NAS Deep Learning Model for Visually Impaired People

Chhabra Payal, Goyal Sonali

doi:10.5109/7236846

Real-time application of object detections are common, and the area of computer vision dramatically benefits from them. Recognizing grocery items poses a more significant challenge for blind individuals compared to those with normal vision. For that purpose, an effective model HYOLO-NAS, is used to detect groceries to aid the visually impaired by seamlessly converting text to audio messages. In the proposed work, Neural Architecture Search technology is used to dynamically update the weights that design child neural networks with the highest accuracy. The hyperparameter tuning on the child network involves adjusting the learning rate, number of epochs, and L2 regularization of weight decay with an Adaptive Moment Estimation optimizer. Google’s Text-to-Speech (gTTS) transforms text into speech signals. After doing many inference experiments, the Hypertuned YOLO-NAS grocery detection model is introduced. The experimental results show that optimized HYOLO-NAS outperforms various detection algorithms with mAP0.5 reaching 96.80% on Grozi-120 and 97.61% on the Retail Product dataset.

An Efficient Grocery Detection System Using HYOLO-NAS Deep Learning Model for Visually Impaired People

Search this article

Description

Journal

Keywords

Details 詳細情報について

Export

Report a problem