Julia Zimpel
Full-Time Data Analytics 2020-08, Campus Berlin, 09 Oct. 2020
Repository: (https://github.com/Yulizzz123/final_project)
In today's fast-moving world of quick real-life encounters, fashion becomes the more crucial in presenting oneself, and leaving a lasting impression on other people. The situation has been hightened by Covid-19 where leaving the home has become severely restricted, and dressing up is a rare chance to 'make it count'.
At the same time, general interest in street-fashion over Haute Couture rises. At the end of 2019, Amazon released the app 'StyleSnap' which allows people to take pictures of favored fashion pieces, and receive purchase recommendations in the online shop.
This project seeks to develop a model that recommends fashion for women after having classified an input image according to clothing type and color. The following rules are applied:
| Input Image | Recommended |
|---|---|
| Top | Skirt |
| Skirt | Top |
| Dress | Dress |
Since many women do not wish to be suspected of wearing the same clothes on two subsequent days (unpublished survey), a fashion piece of a different color is recommended.
Recommenders may benefit several parties. Companies may increase revenue and customer satisfaction through cross-selling, while customers may save time in purchasing multiple items at once as well as in finding favored choices. Hence, this project is also a contribution to the study of business-customer interactions.
This project is mainly written in Python. For image recognition the state-of-the-art method of Convolutional Neural Networks (CNN) is employed by using the Tensorflow. CNN is a method from Deep Learning which has a higher efficiency and accuracy than other image recognition methods, such as kNN. For color detection, kNN is sufficient, and hence employed.
Two workbook series (A and B) are created with a sample size of 12,000 (A) and 1,500 (B) images. While series A achieves a higher model accuracy in image recognition, series B allows for higher flexibility in adjusting statements and parameters due to its smaller requirement on processing capacity. Hence, B was created before moving to the larger sample of A.
Working with large data volumes has led to the application of following insights for faster processing:
| No. | Principles |
|---|---|
| 1. | Images are shrunk in size for deep learning. |
| 2. | Rather than creating long functions, code is split up into small components, and batches are employed. |
| 3. | To avoid long reruns of code, modules and separate worksheets are created. |
The model is created with 6 convolutional and pool layers and pools to which the input sequentially passes.
Adjusting the hyperparameters, the following has proven to achieve an optimal result with the data at hand:
| No. | Findings |
|---|---|
| 1. | 6 convolutional and pool layers rather than smaller and higher numbers of layers |
| 2. | 96 filters rather than a mix of 32, 64 and 96 filters |
| 3. | 512 neurons rather than 128 neurons |
| 4. | Adding a dropout layer |
| 5. | A dropout layer of 70% rather than 50% or 90% |
| 6. | 15 epochs rather than 10, 12 or 20 |
The large sample size of series A with 12,000 images achieves a test accuracy of 72%, while the smaller sample size of series B with 1,500 images achieves a test accuracy of 65%. The difference of only 7 percentage points lower of test accuracy of an eight times smaller sample size suggests that by choosing the right hyperparamters already considerable results can be obtained.
The following sample datasets are used in the project:
| Category | Number of Images | A Selected Images | B Selected Images |
|---|---|---|---|
| Tops | 10,078 | 4,000 | 500 |
| Skirt | 12,742 | 4,000 | 500 |
| Dress | 60,768 | 4.000 | 500 |
Since the samples are considerable low, only 10% of A and B respectively are for testing to reserve a maximum amount of images for training.
The DeepFashion database contains over 800,000 images. From this the attribute prediction subset of 290,000 images is selected, of which further subsets for training and testing are formed (see dataset).
"DeepFashion" (http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion.html) "Attribute Prediction" (http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion/AttributePrediction.html)
For testing with a different database, the Fashion Product Image database is used, which comprises 44,440 images. "Fashion Product Images" (https://www.kaggle.com/paramaggarwal/fashion-product-images-dataset)
| No. | Activity |
|---|---|
| 1. | Clean Data |
| 2. | Develop Models for Image Recognition |
| 3. | Test Models for Image Recognition |
| 4. | Develop Model for Color Recognition |
| 5. | Issue Recommendation |
For the project a series A and B created with different sample sizes (see 2. Methodology). Workbooks AB refer to files that concern both series A and B, i.e. images from a different dataset are prepared for the model to be applied on.
The project folder contains the following files:
| No. | File | Description |
|---|---|---|
| 1. | README | - summary of project outline |
| 2. | A1_Data_Preprocessing.ipynb | - prepare data for series A |
| 3. | A2_Extract_Transform.ipynb | - transform data for series A |
| 4. | A3_Create_and_Test_Model.ipynb | - create and test model for series A |
| 5. | A4_Recommendations.ipynb | - create cross-selling offers |
| 6. | B1_Data_Preprocessing.ipynb | - prepare data for series B |
| 7. | B2_Extract_Transform.ipynb | - transform data for series B |
| 8. | B3_Create_and_Test_Model.ipynb | - create and test the model for series B |
| 9. | B4_Recommendations.ipynb | - create cross-selling offers |
| 10. | AB1_Data_Wrangling_New_Test.ipynb | - prepare different dataset for simulation |
| 11. | AB2_Extract.ipynb | - prepare data access |
| 12. | style_data.csv | - selection of images for the AB series |
| 13. | styles.csv | - original data of the AB series |
| 14. | a_data.pickle | - load data for series A |
| 15. | b_data.pickle | - load data for series B |
| 16. | a_recognition.h5 | - model of series A |
| 17. | b_recognition.h5 | - model of series B |
Next I will increase the test accuracy of my model by adjusting further hyperparameters and the model's architecture. Moreover, I will incorporate a recommendation engine based on deep learning in this project.
The following sources have been used in this project:
Image Recognition in Python with TensorFlow and Keras (https://stackabuse.com/image-recognition-in-python-with-tensorflow-and-keras/)
Flower Classification with Deep Neural Network with Tensorflow and Python Programming (https://www.youtube.com/watch?v=POO1gdUJ7yE)
Classify Images Using Convolutional Neural Networks (https://medium.com/@randerson112358/classify-images-using-convolutional-neural-networks-python-a89cecc8c679)
Deep Transfer Learning for Image Classification (https://towardsdatascience.com/deep-transfer-learning-for-image-classification-f3c7e0ec1a14)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent