Dynamic Task Queue for Retail Catalog Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High latency and inefficiency in processing large datasets of images using multiple inference models for tasks such as non-compliant content detection in retail catalogs, requiring significant computational resources and human intervention.
Innovation Solution
A system utilizing a task populator and distributed task queue with CPU and GPU clusters, enabling efficient processing of large image sets by dynamically assigning evaluation models and distributing tasks across multiple processors, and utilizing a durable messaging system for data exchange between processors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple inference models are applied to process images in retail catalog, then classification accuracy and content detection capability are improved, but processing time and computational resource requirements increase significantly
Solution Approach 1:
The patent segments the processing workflow by dividing images into different categories (e.g., clothing, footwear, accessories) and applying specific inference models only to relevant segments. This selective application of models based on image segmentation reduces overall processing time while maintaining classification accuracy for each category.
Solution Approach 2:
The system performs preliminary actions by first applying a lightweight classification model to categorize images before applying more complex inference models. This preliminary sorting reduces the computational burden on subsequent models by pre-filtering images that require detailed analysis, thereby reducing processing time while preserving accuracy.
2Measurement precision
If multiple inference models are applied to process images in retail catalog, then classification accuracy and content detection capability are improved, but computational resource requirements increase significantly
Solution Approach 1:
The patent segments the processing workload by dividing images into different categories and applying specific inference models only to relevant segments. This selective application of models based on image segmentation reduces overall computational resource requirements while maintaining classification accuracy for each category.
Solution Approach 2:
The system applies partial action by using a tiered approach where only necessary inference models are applied to each image based on its category. Not all images require all models - the system applies the minimum necessary computational resources to achieve accurate classification for each specific image type.
3Adaptability or versatility
If constant human intervention is used to guide processing workflow for each image with respect to each inference model, then processing flexibility and adaptability are improved, but processing efficiency and throughput decrease
Solution Approach 1:
The patent implements self-service by enabling the system to automatically guide its own processing workflow without constant human intervention. The workflow manager autonomously routes images to appropriate inference models based on image metadata and model capabilities, allowing the system to adapt and process images efficiently while maintaining flexibility through automated decision-making.
Data Source
AI summary
Systems and methods including one or more processors and one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform receiving item data for items from a catalog, assigning a task for evaluation of the item data, storing a plurality of task jobs to a task queue, repeatedly setting, in real time, a respective processor to perform a respective evaluation model, processing the plurality of task jobs stored to the task queue by determining, in real time, whether a first evaluation model set to be performed on a first processor is capable of meeting the first evaluation criteria of the first task data, performing, on the first processor, the first evaluation model on the first task data, and transmitting first first-evaluation-model-output instructions, and repeatedly updating, in real time, the task queue. Other embodiments are disclosed herein.