AI
AI
We first scan the scrap pieces into a folder. We then use Intel’s OpenCV to feed the scrap pieces into a convolutional network, specifically Resnet50, which has been trained on images from ImageNet and can detect the desired shapes. Since it's pre-trained, we don't need to initially feed it with pre-existing shapes. As users use our machine, we use a back propagation algorithm to improve detection and segregation of the scraps.
We employ a pop-up with a weighted rating and a textbox for the user to describe the actual image. For working out the fashion sense of a person and recommending clothing pieces based on their style, we use FSNet Recommender. Under user consent, FSNet analyzes their Instagram, Facebook, and Pinterest to recognize and tag each individual style of the person based on the DeepFashion2 Dataset.
That tagged data is stored on a log file, which is then input into another instance of Resnet50. It combines the data from scrap analysis and seeks to fit the scrap pieces inside the garments from the DeepFashion2 database. It also works based on a CLO3D compatible template, which is the DXF-AAMA filetype as it contains texture information and is 2D.
After combining the scrap pieces into garments, it gives a limited number of outputs based on the percentage of how much the scrap pieces match. A web app accompanies this, which connects to a backend python script and a CLO Viewer web app. The CLO Viewer allows us to visualize the DXF file in 3D. The web app also has a user interface where one can input their social media accounts, height, and body dimensions. This information is stored in an XML file, compiled in a compatible manner for CLO3D.
Once a choice has been made, the user can continue and print the blueprints or edit the garments themselves on their computer through CLO3D for more advanced users. Once an advanced user has finished modifying the garment in the software, they can upload the file back to the Fabrique via a Type-C Cable or cloud-based synchronization. Once the file has been imported into Fabrique, it is read with a bash shell script and input into Browzwear.
If the user has not modified the garment, it is automatically sent into Browzwear as well. Browzwear allows us to generate stitching patterns into a PSD compatible format, which easily allows us to overlay an A4 ISO standard using a simple script. PSD files can also be batch scripted into a PDF file. This file is then handled by Incron, a daemon and table manipulator, which takes the job of printing the file.
Possible Improvements
We could use various techniques to preprocess the data before feeding it into the system, as the quality of the input data significantly affects the accuracy of the system's output. For example, we could utilize image enhancement techniques, such as denoising and contrast adjustment, to improve the quality of the images. Moreover, we could augment the dataset to enhance the model's ability to generalize to new inputs.
When it comes to feature engineering, we could transform raw data into meaningful features that could be used to train the machine learning model. This approach can help us reduce the dimensionality of the problem and improve the model's performance. For instance, we could extract relevant features from scrap pieces, such as their color, texture, and shape, and combine them with the user's style preferences to generate garment recommendations.
In terms of model architecture, we could experiment with different state-of-the-art models, such as EfficientNet or DenseNet, to achieve better accuracy with less computational cost. Additionally, we could consider using an ensemble of models to enhance the system's robustness.
To optimize the system's performance, we could consider hardware optimization techniques such as using a GPU instead of a CPU to accelerate the model training and inference. We could also utilize distributed computing to parallelize the computations and reduce the processing time.