Object Detection Neural Network Computational Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based object detection processes are computationally intensive, with a significant portion of the execution time spent on the detection layer, which performs complex operations on anchor boxes, scales, and object classifications.
Innovation Solution
The method involves using a convolution neural network to determine image classifications and confidence levels, calculating upper-bound scores based on these confidence levels, and performing activation functions on these scores to reduce the computational complexity of the detection layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional detection layer operations are performed for each anchor box, scale, and object classification, then complete object detection is achieved, but processing time and computational complexity increase significantly
Solution Approach 1:
The patent extracts and performs critical detection operations (scoring, activation functions, decoding) on a CPU before transferring data to the DSP. This separates time-critical operations from compute-intensive operations, allowing the DSP to focus only on essential detection tasks and reducing overall processing time while maintaining detection accuracy
Solution Approach 2:
The patent performs preliminary processing of detection layer operations on the CPU before data transfer to the DSP. By completing scoring, activation functions, and decoding in advance, the system reduces the computational burden on the DSP and minimizes processing time for the critical detection path
2Use of energy by moving object
If detection layer operations are executed on DSP with lesser compute capability, then power consumption is reduced, but processing speed decreases
Solution Approach 1:
The patent segments the detection layer operations into two parts: time-critical operations (scoring, activation functions, decoding) executed on the CPU, and essential detection operations executed on the DSP. This segmentation allows each processor to operate in its optimal performance range, with the CPU handling operations that benefit from higher speed and the DSP handling operations that benefit from lower power consumption
3Measurement precision
If compute-intensive operations are performed at activation layers for each anchor box and scale, then complete classification is achieved, but the number of operations increases
Solution Approach 1:
The patent performs activation functions and decoding operations only on a subset of anchor boxes and scales that are most likely to contain objects, rather than processing all possible combinations. This partial action approach maintains classification precision for relevant objects while significantly reducing the total number of operations required
Data Source
AI summary
Various embodiments of the present technology relate to using neural networks to detect objects in images. More specifically, some embodiments relate to the reduction of computational analysis regarding object detection via neural networks. In an embodiment, a method of performing object detection is provided. The method comprises determining, via a convolution neural network, at least a classification of an image, wherein the classification corresponds to an object in the image and comprises location vectors corresponding to pixels of the image. The method also comprises, for at least a location vector of the location vectors, obtaining a confidence level, wherein the confidence level represents a probability of the object being present at the location vector, and calculating an upper-bound score based at least on the confidence level. The method further comprises, for at least an upper-bound score based at least on the confidence level, performing an activation function on the upper-bound score, and classifying, via a detection layer, the object in the image.