Neural Network Quantization Control for Saturation-Aware Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection methods suffer from degradation in detection accuracy due to an imbalance between the number of saturation layers and average saturations, as the threshold for decimal point position change is uniform across all input images, leading to increased saturation layers and reduced accuracy.
Innovation Solution
An object detection device and method that dynamically adjusts the decimal point position for each layer in a multilayer neural network based on the detection results of multiple input images, optimizing the upper and lower limit saturation thresholds to minimize saturation layers and maintain accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a uniform threshold is used for decimal point position change across all input images, then the device complexity is reduced and operation is simplified, but the number of saturation layers increases and detection accuracy degrades
Solution Approach 1:
The patent applies dynamics by transitioning from a static uniform threshold to a dynamic adaptive threshold that changes based on input image characteristics. The threshold determination unit calculates saturation thresholds dynamically for each input image based on statistical information from multiple images, allowing the system to adapt to varying input conditions and maintain detection accuracy while managing device complexity.
Solution Approach 2:
The patent implements parameter changes by adjusting the threshold parameter based on input image characteristics. Instead of using a fixed threshold value, the system changes the threshold parameter dynamically according to the statistical information derived from multiple input images, thereby optimizing detection accuracy for different input conditions without significantly increasing device complexity.
2Use of energy by stationary object
If the data bit width is reduced to 8-16 bits for size reduction and lower power consumption, then circuit scale and power consumption are reduced, but saturation layers increase and detection accuracy degrades
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the threshold parameter based on input image characteristics. This allows the system to maintain optimal detection accuracy with reduced bit width (8-16 bits) by adapting the threshold to each input image's statistical properties, thereby preventing excessive saturation layers while consuming less power.
Solution Approach 2:
The patent implements feedback by using the saturation threshold calculation results to guide subsequent processing decisions. The threshold determination unit continuously monitors saturation conditions and adjusts processing parameters accordingly, creating a feedback loop that maintains detection accuracy despite reduced bit width and lower power consumption.
3Productivity
If the decimal point position is changed to reduce average saturations, then processing efficiency may improve, but the number of saturation layers increases and detection accuracy degrades
Solution Approach 1:
The patent applies dynamics by making the threshold determination a dynamic process that responds to input image characteristics. Rather than statically changing decimal point positions to reduce saturations, the system dynamically calculates appropriate thresholds for each input image, maintaining detection accuracy while adapting to varying processing conditions and efficiency requirements.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An object detection device subjects fixed-length data having a decimal point position set therein to an arithmetic processing corresponding to respective layers in a plurality of layers cofiguring a multilayer neural network to which an input image is input, the arithmetic processing being performed in accordance with a processing algorithm for the multilayer neural network to which an input image is input. In the arithmetic processing, the object detection device counts the upper limit number of saturations, which is a number of times that upper limit value of a value range determined by the decimal point position is exceeded, and the lower limit number of saturations, which is a number of times that the lower limit value of the value range is not reached. The object detection device counts the upper limit number of saturation layers, which is a number of layers in which the upper limit number of saturations is one or larger, and the lower limit number of saturation layers, which is a number of layers in which the lower limit number of saturations is one or larger. The object detection device changes at least one of the upper limit saturation threshold, which is the threshold of the upper limit number of saturations or the lower limit saturation threshold, which is the threshold of the lower limit number of saturations, when at least one of the upper limit saturation threshold or the lower limit saturation threshold is determined not to be optimal based on an amount of change in the upper limit number of saturation layers and an amount of change in the lower limit number of saturation layers. The object detection device sets the decimal point position for each layer in the plurality of layers, based on a result of the determination.