BikeVision
About the project
This project presents a novel approach for predicting multiple design parameters of products using 2D images and supervised learning techniques. The primary goal is to develop an accurate and efficient predictive model for various design parameters, streamlining the design process and enabling designers to explore a wider range of possibilities. The proposed method demonstrates high prediction accuracy and outperforms existing methods for predicting multiple product design parameters.
Method
The proposed method utilizes a pre-trained feature extracting network (ResNet50 CNN) fine-tuned on a target dataset of input images and corresponding output labels. The model architecture includes a series of fully connected layers to perform multilabel prediction for binary and continuous label design parameters.
Data Preprocessing
Images are resized to a fixed size (224x224) and converted to grayscale if necessary. The label data is split into binary and continuous values, allowing the model to learn from both categorical and numerical data simultaneously.
The proposed approach achieves high prediction accuracy for binary parameters (96.42% binary accuracy) and a low error rate for continuous variables (mean squared error of 0.0066). The GradCAM values reveal the importance of different features in the model’s predictions, suggesting that the rear of the bicycle and the frame carry more weight in the predictions than the front.
Results
Project Presented at IDETC 2023! Read It Here!
Parameter Extraction from Images using Multilabel Supervised Learning
Findings
The research demonstrates the effectiveness of using 2D images and design parameters as labels to train a supervised network for predicting design parameters of products. The high accuracy achieved for binary parameters and low error rate for continuous variables indicate the potential of this approach as a reliable tool for predicting multiple design parameters in practical engineering design scenarios, ultimately accelerating the design process and enabling designers to explore a wider range of design possibilities.
Let’s talk?
If you have any questions about my qualifications or would like to discuss further, please don't hesitate to get in touch!
