FPGA Scene Text Recognition With Rotated Text Box Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing methods, particularly in scene text recognition, face challenges with low efficiency due to serial processing by CPUs and high costs and power consumption of GPUs, and existing FPGA implementations do not effectively address parallel processing of large datasets.
Innovation Solution
A task data processing method utilizing FPGA units to perform subtasks in parallel for multiple pieces of task data, allowing concurrent execution of substructures in a machine learning model, with CPU and FPGA units working together to improve efficiency and reduce power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing methods are used to ensure accurate recognition and processing of visual information, then processing accuracy is maintained, but processing speed becomes slow and cannot meet real-time requirements
Solution Approach 1:
The patent divides image processing into two distinct stages: coarse processing using a lightweight network for rapid initial results, and fine processing using a sophisticated network for accurate refinement. This segmentation allows the system to achieve both speed (through the fast coarse stage) and accuracy (through the precise fine stage) by distributing different functional requirements to specialized processing components.
2Measurement precision
If complex algorithms are deployed to improve recognition accuracy, then processing precision increases, but device complexity and computational resources increase
Solution Approach 1:
The complex recognition task is segmented into two processing stages with different complexity levels. The first stage uses a simplified lightweight network that handles basic recognition quickly, while the second stage employs a more complex sophisticated network only when needed for refinement. This reduces overall device complexity compared to using the sophisticated network for all processing.
Solution Approach 2:
Instead of always applying the full complex algorithm, the system applies a simplified version (lightweight network) for most cases and only activates the more complex processing (sophisticated network) when refinement is needed. This partial application of complexity achieves high accuracy where required while maintaining low complexity for routine operations.
3Speed
If a single processing model is used to balance speed and accuracy, then system simplicity is maintained, but both speed and accuracy cannot be optimized simultaneously
Solution Approach 1:
The system dynamically adjusts processing intensity by switching between two processing modes: a fast mode using the lightweight network for speed-critical operations, and a precise mode using the sophisticated network for accuracy-critical operations. This dynamic adaptation allows the system to optimize performance characteristics based on real-time requirements rather than being constrained by a fixed single-model approach.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present application relates to an image processing method, and a task data processing method and device. The image processing method, implemented by using processing units corresponding to sub-structures in a machine learning model to execute corresponding sub-tasks, wherein at least some of the processing units comprise FPGA units, comprises: obtaining an image to be processed; extracting a text feature in the image to be processed; determining candidate text boxes of any angles in the image to be processed according to the text feature; performing pooling processing on the candidate text boxes to rotate a region of interest, and projecting the candidate text boxes onto a feature map of a fixed size to obtain a text box feature map corresponding to the candidate text boxes; and recognizing text in the text box feature map to obtain a text recognition result. According to the present application, data can be processed in parallel by means of an FPGA architecture to realize the image processing method, costs and power consumption are reduced, and the accuracy and efficiency of text recognition in images to be processed are improved.