KR-20260064638-A - NEURAL NETWORK FEATURE MAP ENCODING/DECODING METHOD, DEVICE, AND RECORDING MEDIUM

Abstract

A neural network feature map encoding/decoding method, device, and recording medium may include a step of acquiring feature nonlinear transformation information, a feature nonlinear transformation step of nonlinearly transforming a feature map based on the feature nonlinear transformation information, and a scaling step of scaling the transformed feature map.

Inventors

• 정순흥
• 이진영
• 최해철
• 한희지
• 이성은

Assignees

• 한국전자통신연구원
• 국립한밭대학교 산학협력단

Dates

Publication Date

20260507

Application Date

20251031

Priority Date

20241031

Claims (1)

1. Step of acquiring feature nonlinear transformation information; A feature nonlinear transformation step that nonlinearly transforms a feature map based on feature nonlinear transformation information; and A neural network feature map decoding method characterized by including a scaling step for scaling the transformed feature map.

Description

Neural Network Feature Map Encoding/Decoding Method, Device, and Recording Medium The present disclosure can be utilized in the technical field of artificial neural networks and signal processing-based encoding/decoding. Video compression technology has evolved to ensure that when a compressed image is restored, the restored image resembles the original as closely as possible, based on human vision. In other words, video compression technology has developed in a direction that minimizes the bit size while simultaneously maximizing the image quality of the restored video. For example, an encoder receives a video as input and generates a bitstream through transformation and entropy encoding processes, while a decoder receives the bitstream and can restore it back into an image similar to the original. Meanwhile, as machine vision performance improves, instances where machines view and consume images instead of humans are increasing. For example, in fields such as smart cities, autonomous vehicles, and airport surveillance cameras, machine-centric image consumption is on the rise. Consequently, interest in image compression technologies centered on machine vision has recently been growing, in addition to human-centric compression. Figure 1 illustrates an example of a convolution operation of multi-channel data that outputs a feature map (2D Array). Figure 2 illustrates an example of a convolution operation that outputs a feature map (3D Array). Figure 3 illustrates examples of a single feature map and a packed feature map. Figure 4 illustrates an example of a feature map encoding structure. Figure 5 illustrates examples of feature map encoding structures, feature nonlinear transformation locations, and feature nonlinear inverse transformation locations. Figure 6 illustrates an example of a feature nonlinear transformation step. Figure 7 illustrates an example of a feature nonlinear inverse transform step. Figure 8 illustrates the feature quantization step. Figure 9 illustrates an example of a method including the start and end values of uniform quantization. Figure 10 illustrates an example of a method including the start and end values of FCTM uniform quantization. FIG. 11 illustrates an example of a method including the start and end values of logarithmic non-uniform quantization. FIG. 12 illustrates an example of a method including the start and end values of inverse logarithmic scale non-uniform quantization. Figure 13 illustrates an example of a method including the start and end values of central symmetric non-uniform quantization. Figure 14 illustrates an example of a method including the start and end values of non-uniform quantization of a normal distribution. Figure 15 illustrates an example of a quantization index mapping step. Figure 16 illustrates an example of a characteristic inverse quantization step. Figure 17 illustrates examples of representative values of the quantization intervals of the quantization index inverse mapping step. The present disclosure is subject to various modifications and may have various embodiments, and specific embodiments may be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present disclosure to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the present disclosure. Similar reference numerals in the drawings may refer to the same or similar functions across various aspects. The shapes and sizes of elements in the drawings may be exaggerated for clearer explanation. For the detailed description of exemplary embodiments described below, reference may be made to the accompanying drawings which illustrate specific embodiments. These embodiments may be described in sufficient detail to enable those skilled in the art to practice the embodiments. It should be understood that various embodiments are different but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the present disclosure in relation to one embodiment. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the embodiment. Accordingly, the following detailed description is not intended to be taken in a limiting sense, and the scope of the exemplary embodiments may be limited only by the appended claims, including all equivalents to those claimed therein, provided they are appropriately described. In this disclosure, terms such as first, second, etc. may be used to describe various components, but said components may not be limited by said terms. Such terms may be used solely for the purpose of distinguishing one component from another. For example, w