【YOLOv8改进 - 注意力机制】NAM:基于归一化的注意力模块,将权重稀疏惩罚应用于注意力机制中,提高效率性能

YOLO目标检测创新改进与实战案例专栏

专栏目录: YOLO有效改进系列及项目实战目录 包含卷积,主干 注意力,检测头等创新机制 以及 各种目标检测分割项目实战案例

专栏链接: YOLO基础解析+创新改进+实战案例

介绍

image-20240630171637321

摘要

识别较不显著的特征是模型压缩的关键。然而,这在革命性的注意力机制中尚未被研究。在这项工作中,我们提出了一种新颖的基于归一化的注意力模块(NAM),该模块抑制了较不显著的权重。它对注意力模块施加了权重稀疏惩罚,从而使其在保留相似性能的同时变得更具计算效率。在Resnet和Mobilenet上与其他三种注意力机制的比较表明,我们的方法可以带来更高的准确性。本文的代码可以在https://github.com/Christian-lyc/NAM公开获取。

文章链接

论文地址:论文地址

代码地址:代码地址

基本原理

NAM (Normalization-based Attention Module)是一种新颖的注意力机制,旨在通过抑制不太显著的特征来提高模型的效率。NAM模块将权重稀疏惩罚应用于注意力机制中,以提高计算效率同时保持性能。NAM模块通过批量归一化(Batch Normalization)的缩放因子来衡量通道的重要性,避免了SE(Squeeze-and-Excitation)、BAM(Bottleneck Attention Module)和CBAM(Convolutional Block Attention Module)中使用的全连接和卷积层。这使得NAM成为一种高效的注意力机制。

NAM模块结合了通道注意力和空间注意力的子模块,利用批量归一化的缩放因子来衡量通道和像素的重要性,从而实现对特征的有效识别和利用。

  1. 通道注意力子模块:使用批量归一化的缩放因子来衡量通道的重要性,通过计算权重来获得输出特征。

image-20240630171918385

  1. 空间注意力子模块:应用像

A ResNet-based Convolutional Decoder-Encoder is a type of neural network architecture that combines the principles of Residual Networks (ResNets) and Decoder-Encoder networks. ResNets are deep neural networks that use skip connections to avoid the vanishing gradient problem and allow for the training of very deep networks. Decoder-Encoder networks, on the other hand, are used for tasks such as image segmentation, object detection, and image generation. The ResNet-based Convolutional Decoder-Encoder architecture consists of a series of encoder layers that downsample the input image and a series of decoder layers that upsample the encoded features to generate the output image. The encoder layers typically consist of Convolutional Layers followed by Batch Normalization and ReLU activation. The decoder layers consist of transposed convolutional layers, also known as deconvolutional layers, followed by Batch Normalization and ReLU activation. The skip connections in the ResNet-based Convolutional Decoder-Encoder architecture allow for the direct transfer of information from the encoder to the decoder layers, which helps to preserve important features and reduce the effects of information loss during the downsampling process. The resulting network can be trained end-to-end using backpropagation to minimize a loss function that measures the difference between the predicted and ground truth images. ResNet-based Convolutional Decoder-Encoder networks have been used successfully for a variety of image reconstruction and generation tasks, including image denoising, super-resolution, and inpainting.
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

YOLO大师

你的打赏,我的动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值