Image Preprocessing Circuit for Neural Network Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems for neural networks require large memory spaces to store whole images, and they either need a digital signal processor to calculate memory addresses for rotated and divided sub-images or cannot perform simultaneous division and rotation of images.
Innovation Solution
An image recognition system and preprocessing method that simultaneously divide and rotate images, classifying each pixel based on a division criterion and rotation angle, and calculating memory addresses for storing sub-images sequentially in memory, thereby reducing memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the whole image is stored in memory before division and rotation, then the image recognition system can retrieve sub-images for processing, but a large memory space is required to store the whole image
Solution Approach 1:
The image is divided into multiple sub-images during the preprocessing stage, and each sub-image is processed separately. This segmentation allows the system to work with smaller data units, reducing the memory space needed compared to storing and processing the entire image at once.
Solution Approach 2:
The image is pre-processed (divided and rotated) before being stored in memory. By performing division and rotation operations during the preprocessing stage rather than during retrieval, the system eliminates the need to store the entire original image, thereby reducing memory requirements while ensuring reliable sub-image retrieval.
2Reliability
If a digital signal processor is used to calculate memory addresses for rotated and divided sub-images, then the image recognition system can retrieve sub-images, but the cost of the digital signal processor is added to the system
Solution Approach 1:
The division and rotation operations are merged into a single preprocessing step that occurs before memory storage. This combined operation eliminates the need for separate address calculation operations that would require a digital signal processor, thereby reducing system complexity and cost while maintaining the ability to retrieve sub-images.
Solution Approach 2:
The preprocessing circuit performs both division and rotation operations autonomously during the preprocessing stage, without requiring external assistance from a digital signal processor. This self-service approach to image processing reduces system complexity by eliminating the need for additional processing components.
3Ease of operation
If the image is rotated through a first layer preprocessing circuit and stored, then the image can be retrieved and divided by a second layer preprocessing circuit, but a large memory space is still required and simultaneous division and rotation cannot be performed
Solution Approach 1:
The division and rotation operations are merged into a single integrated preprocessing step. Instead of performing rotation in one circuit and division in another separate circuit at different stages, both operations are performed simultaneously during the preprocessing phase, eliminating the need to store the entire rotated image and reducing memory requirements.
Solution Approach 2:
Both division and rotation operations are performed as preliminary actions during the preprocessing stage before the image data is stored in memory. This approach allows the system to work with already-processed sub-image data during retrieval, eliminating the need for large memory space while maintaining ease of operation.
Data Source
AI summary
An image recognition system for a neural network and an image preprocessing method are provided. The image recognition system includes a memory, an image sensor, a preprocessing circuit, and a neural network processing member. The image sensor is configured to obtain an image that includes a plurality of pixels. The preprocessing is configured to classify each pixel based on the division criterion and the rotation angle as belonging to at least one of the target sub-images, and calculate a memory address corresponding to the pixel in the target sub-image to which the pixel belongs, so as to sequentially store the plurality of target sub-images in the memory. The neural network processing member is configured to retrieve one of the stored target sub-images from the memory for recognition.