Building a Product Image Classifier
Data Collection:
◦ Snapshots were taken from different categories in the Slash app.
◦ Images were split into 4 categories:
▪ 0 => Nutrition
▪ 1 => Fashion
▪ 2 => Games
▪ 3 => Accessories
◦ These categories were chosen as a proof of concept and for ease of expanding the model to handle more categories. They also have high-quality images requiring less preprocessing.
• Choosing the Model:
◦ Due to the small dataset, training a model from scratch was not feasible.
◦ Pretrained models were considered, such as ResNet, MobileNet, and VGG.
◦ VGG16 was selected as it showed the best performance for the use case.
◦ The weights were frozen, and only the last layer was fine-tuned using softmax.
◦ Training accuracy was 100%, but test accuracy was around 92%, indicating some overfitting. This can be addressed by adding more training examples.
• Save the Model:
◦ The trained model was saved for future inference.
• Inference on Different Images:
◦ Inference was performed on four different images, and the results are as follows:
