KR-102963257-B1 - IMPROVEMENT OF DETECTION ACCURACY-WORKLOAD TRADEOFF WITH INVOLUTION AND FILTER PRUNING IN AUTONOMOUS VEHICLE ENVIRONMENT

KR102963257B1KR 102963257 B1KR102963257 B1KR 102963257B1KR-102963257-B1

Abstract

A technology for improving the detection accuracy-workload trade-off by applying involution and filter pruning in an autonomous vehicle environment is disclosed. A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment, performed by a trade-off improvement system according to one embodiment, may include the steps of: constructing a new object detection model by applying a filter pruning technique to an object detection model; obtaining an integrated performance result of the newly constructed object detection model through an integrated performance indicator provided to measure the resolution of the trade-off between detection accuracy and workload; and reconstructing the new object detection model by applying an involution technique to the newly constructed object detection model according to the obtained integrated performance result.

Inventors

정용화
노재현
천수
백화평
유승현

Assignees

고려대학교 세종산학협력단

Dates

Publication Date: 20260511
Application Date: 20230912

Claims (13)

A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment performed by a trade-off improvement system, A step of configuring a second object detection model by applying a filter pruning technique to a first object detection model; A step of obtaining the integrated performance results of the second object detection model configured above through integrated performance indicators provided to measure the resolution of the trade-off between detection accuracy and workload; and A step of reconstructing a third object detection model by applying an involution technique to the second object detection model configured above according to the integrated performance results obtained above. A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment including
In paragraph 1, A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment, characterized in that the first object detection model is a Convolutional Neural Network-based YOLOv7-Tiny model.
In paragraph 1, The step of configuring the above-mentioned second object detection model is, A step of pruning a filter of a target layer set in the first object detection model or a stage separated in the first object detection model. A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment including
In paragraph 1, The step of obtaining the above integrated performance results is, If the integrated performance result and average accuracy of the second object detection model configured above are increased compared to the integrated performance result and average accuracy of the first object detection model, the pruning ratio is increased, and according to the increased pruning ratio, the filter of the target layer set in the second object detection model configured above or the filter of the stage classified in the second object detection model configured above is pruned. A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment including
In paragraph 1, The step of reconstructing the above-mentioned third object detection model is, If the integrated performance result and average accuracy of the second object detection model configured above are not increased compared to the integrated performance result and average accuracy of the first object detection model, a step of replacing some convolution layers of the second object detection model with involution layers. A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment including
In paragraph 1, The step of configuring the above-mentioned second object detection model is, A step of acquiring image information using a camera installed in an autonomous vehicle; and The step of inputting the above-mentioned acquired image information into the first object detection model A method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment including
A computer program stored on a non-transient computer-readable recording medium for executing a method for improving the trade-off between detection accuracy and workload in an autonomous vehicle environment according to any one of claims 1 to 6 on the trade-off improvement system.
In the trade-off improvement system, A lightweight model component that configures a second object detection model by applying a filter pruning technique to a first object detection model; A performance result acquisition unit that obtains the integrated performance result of the second object detection model configured above through an integrated performance indicator provided to measure the resolution of the trade-off between detection accuracy and workload; and A lightweight model reconstruction unit that reconstructs a third object detection model by applying an involution technique to the second object detection model configured above according to the integrated performance results obtained above. A trade-off improvement system including
In paragraph 8, A trade-off improvement system characterized in that the first object detection model is a Convolutional Neural Network-based YOLOv7-Tiny model.
In paragraph 8, The above lightweight model components are, Pruning the filter of the target layer set in the first object detection model or the stage separated in the first object detection model A trade-off improvement system characterized by
In paragraph 8, The above performance result acquisition unit is, If the integrated performance result and average accuracy of the second object detection model configured above are increased compared to the integrated performance result and average accuracy of the first object detection model, the pruning rate is increased, and according to the increased pruning rate, the filter of the target layer set in the second object detection model configured above or the filter of the stage separated in the second object detection model configured above is pruned. A trade-off improvement system characterized by
In paragraph 8, The above lightweight model reconstruction unit is, If the integrated performance result and average accuracy of the second object detection model configured above are not increased compared to the integrated performance result and average accuracy of the first object detection model, some convolution layers of the second object detection model are replaced with involution layers. A trade-off improvement system characterized by
In paragraph 8, The above lightweight model components are, Acquiring image information using a camera installed in an autonomous vehicle and inputting the acquired image information into the first object detection model A trade-off improvement system characterized by

Description

Improvement of Detection Accuracy-Workload Tradeoff with Involution and Filter Pruning in an Autonomous Vehicle Environment The following description relates to a technology for improving the trade-off between accuracy and workload for deep learning-based object detection in an autonomous vehicle environment. With the recent advancement of deep learning technology, deep learning techniques are being applied to various image processing applications; however, to apply them to actual object detection applications in autonomous vehicle environments, a trade-off between accuracy and workload must be considered. If experiments are conducted intensively to improve accuracy without considering time in a fast-moving autonomous driving environment, real-time detection may not be satisfied in the actual environment, potentially leading to serious accidents. Figure 1 is a diagram illustrating a convolution operation method. FIG. 2 is a diagram illustrating an involution operation method in one embodiment. FIG. 3 is a diagram illustrating the operation of acquiring image information in an autonomous driving environment vehicle in one embodiment. FIG. 4 is a diagram showing the stages in the object detection model structure in one embodiment. FIG. 5 is a diagram illustrating the operation of pruning a filter of a specific layer in an object detection model in one embodiment. FIG. 6 is a diagram illustrating the operation of replacing a convolution layer with an involution layer in stage 5 of an object detection model in one embodiment. FIG. 7 is a block diagram illustrating the configuration of a trade-off improvement system in one embodiment. FIG. 8 is a flowchart illustrating a method to improve the trade-off between detection accuracy and workload in an autonomous vehicle environment in one embodiment. Hereinafter, embodiments will be described in detail with reference to the attached drawings. In the embodiments, we will describe an operation that improves the trade-off between detection accuracy and workload by applying involution and filter pruning techniques to an object detection model. Figure 1 is a diagram illustrating a convolution operation method. Convolutional computation uses a method that performs operations separately for each channel, regardless of position within the channel. Convolutional computation possesses Channel-Specific and Spatial-Agnostic characteristics. Due to these characteristics, a new kernel is used for each channel, and since the same kernel is applied to a channel regardless of pixel position, kernels are reused to reduce the number of parameters and extract specific features per kernel. However, because it is created by synthesizing information from the previous layer for each channel, the redundancy of kernels increases as the number of channels increases, leading to similar results. Furthermore, since it operates based on kernel size and stride values, it has the disadvantage of being difficult to grasp spatial relationships over a wide range at once. In Figure 1, one of the three kernels of the filter is used entirely within one channel regardless of the position of the pixel within the channel, and the operation of all kernels is added to become the result of one pixel in the output channel. Accordingly, in the embodiment, an involution operation method is used to improve the disadvantages of the convolution operation method. FIG. 2 is a diagram illustrating an involution operation method in one embodiment. Involution possesses Channel-Agnostic and Spatial-Specific characteristics, which are the exact opposite of Convolution. That is, in involution, a different kernel is generated for each location within a channel, and the generated kernel is used across multiple channels. Figure 2 shows how the kernels generated for each location participate in the operation across multiple channels. Compared to Convolution, the involution operation method uses larger kernels and learns 1×1 Convolutions that generate the kernels, rather than learning all the kernels that differ for each location; consequently, this reduces model size and workload while improving accuracy. Specifically, pixel values at the same location are extracted across all channels to generate kernels through 1×1 Convolution. The generated kernels are then applied to the locations where pixel values were extracted for each channel, and the results are added together to obtain the output of a single pixel in each output channel. As such, compared to the convolution method, the involution method can reduce the number of parameters and the amount of computation required for the operation. FIG. 3 is a diagram illustrating the operation of acquiring image information in an autonomous driving environment vehicle in one embodiment. At least one camera may be installed at a specific location (e.g., exterior/interior, top/bottom/left/right) of the autonomous vehicle. The at least one camera installed at a specific location of the au