Image Sensor Multi-Stage AI Inference for Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

There is a demand for further functionality improvement of image sensors capable of performing inference processes using artificial intelligence models.

Innovation Solution

An image sensor with a pixel array unit and an inference processing unit that executes multiple inference processes using different artificial intelligence models based on image data output, enhancing its functionality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single artificial intelligence model is deployed in the image sensor, then the device maintains simplicity, but the functionality and inference capabilities are limited

Engineering Contradiction:

Improveinference capabilitiesVSAvoidmodel structure

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent divides the inference task into multiple stages, each handled by a separate AI model. The first model performs initial inference on raw image data, and the second model performs further inference on the intermediate results. This segmentation allows the system to achieve complex functionality while keeping each individual model relatively simple and manageable.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a temporal dimension to the inference process by executing models sequentially rather than simultaneously. The first inference process generates intermediate results that are then fed into the second inference process, creating a multi-stage pipeline that enhances functionality without requiring all models to operate at once, thus managing complexity.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Adaptability or versatility

If multiple inference processes are executed sequentially, then the functionality is improved, but the processing time increases

Engineering Contradiction:

Improveinference functionalityVSAvoidprocessing time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The first inference process performs preliminary processing on the image data, generating intermediate results that are pre-prepared for the second inference process. This preliminary action allows the second model to work with refined data, potentially reducing its processing time and improving overall efficiency despite the sequential nature of the operations.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent ensures continuous useful action by piping the output of the first inference process directly into the second inference process without idle gaps. The intermediate results are immediately utilized, maintaining a continuous flow of data processing and maximizing the productivity of the multi-model system.

Inventive Principle:

Principle #20Continuity of useful action

Data Source

PatentUS20250292527A1Image sensor, information processing method, and program

Publication Date: 2025.09.18 SONY SEMICON SOLUTIONS CORP

AI summary

An image sensor according to the present technology includes a pixel array unit where multiple pixels are two-dimensionally arrayed, and an inference processing unit that executes a first inference process using a first artificial intelligence model on the basis of image data output from the pixel array unit and that executes a second inference process using a second artificial intelligence model on the basis of a result of the first inference process.