shopping-intent

Google’s Shopping Intent Classifier

Model Overview

This model is a multi-label image classification model extracted from Chrome. It's Google’s custom convolutional neural network in TensorFlow Lite v3 format based on mobilenet V3 small model. It can be deployed in an automated pipeline capable of classifying product images in bulk. The model is designed to determine whether an image is optimized for shopping intent and can classify images into one of four categories related to shopping intent.

Model Details

• Model Name: shopping_intent_x_sensitivity_classifier
• Checkpoint: mobilenet_v3_small_224_04132253_ckpt_3006395
• Description: The model is MLIR Converted and classifies whether an image has shopping intent or is sensitive.
• Model Author: lens-proactive-dev
• Framework: TensorFlow Lite v3
• Architecture: MobileNet V3 Small

Inputs

• name: normalized_input_image_tensor
• tensor: float32[1,224,224,3]
• denotation: Image(RGB)
• Description: Input image to be classified. The input is expected to be an RGB image with type UINT8.
• identifier: 0

Outputs

• name: shopping_intent

• tensor: float32[1,4]

• denotation: Feature

• Description: Probability whether the image has shopping intent

• identifier: 222

• name: sensitive

• tensor: float32[1,2]

• denotation: Feature

• Description: Probability whether the image is sensitive

• identifier: 220

Labels

The model provides classifications across four shopping intent categories:

• LABEL_1: shopping_intent:negative
• LABEL_2: shopping_intent:apparel
• LABEL_3: shopping_intent:home_decor
• LABEL_4: shopping_intent:other

The model provides classifications across two sensitivity categories:

• LABEL_1: sensitive:negative
• LABEL_2: sensitive:positive

Model Use Cases

This model can be used for:

• Determining whether an analyzed image is misclassified by shopping intent.
• Identifying misclassified shopping categories.
• Detecting images with ambiguous intent and category.

How It Works

The model takes in a pre-processed image (224x224) and returns two sets of probabilities:

• Shopping intent (4 labels)
• Sensitive image (2 labels)

Technical Specifications

Model Inputs:

• name: normalized_input_image_tensor
• tensor: float32[1,224,224,3]
• denotation: Image(RGB)
• Description: Input image to be classified. The input is expected to be an RGB image with type UINT8.
• identifier: 0

Model Outputs:

• name: shopping_intent

• tensor: float32[1,4]

• denotation: Feature

• Description: Probability whether the image has shopping intent

• identifier: 222

• name: sensitive

• tensor: float32[1,2]

• denotation: Feature

• Description: Probability whether the image is sensitive

• identifier: 220

Model Architecture

The full model architecture is available as: PNG | SVG

Additional Information

• Original Article: Link to the article

Practical Application

Interested in using this model in an automated pipeline for bulk image classification? Please book an appointment to discuss your needs.