多尺度特征融合下三維視覺圖像場景分割算法

摘 "要： 為減少噪聲對分割結(jié)果的影響，降低單一尺度特征對分割結(jié)果的敏感性，提升分割算法的魯棒性與穩(wěn)定性，并增強分割邊界清晰度，提高分割精度，文中提出一種多尺度特征融合下三維視覺圖像場景分割算法。雙路徑多信息域注意力模塊通過結(jié)合頻域通道與空間注意力機制，提取三維視覺圖像的多尺度特征，降低單一尺度特征對分割結(jié)果的敏感性；在多尺度特征融合模塊內(nèi)添加空洞卷積層，增大多尺度特征的感受野，并融合增大感受野的多尺度特征，捕捉圖像的細節(jié)信息和全局信息，減少噪聲對分割結(jié)果的影響，提升分割算法的魯棒性與穩(wěn)定性；利用Softmax分類器處理融合特征，得到三維視覺圖像場景分割結(jié)果；通過全連接條件隨機場、后處理分割結(jié)果，優(yōu)化分割邊界清晰度，提高分割精度。實驗結(jié)果證明：該算法可有效提取三維視覺圖像的多尺度特征，有效完成三維視覺圖像場景分割，且場景分割的邊界非常清晰。為三維視覺圖像的處理與分析提供了新的思路和方法。

關(guān)鍵詞： 多尺度； 特征融合； 三維視覺； 圖像場景分割； 注意力機制； 空洞卷積； Softmax分類器； 條件隨機場

中圖分類號： TN911.73?34； TP391.41 " " " " " " " 文獻標識碼： A " " " " " " " " " 文章編號： 1004?373X（2024）21?0046?05

3D visual image scene segmentation algorithm based on multi?scale feature fusion

YAN Jingfu， WANG Pengfei

（China University of Petroleum （Beijing） at Karamay， Karamay 834000， China）

Abstract： A 3D visual image scene segmentation algorithm based on multi?scale feature fusion is proposed to reduce the impact of noise on the segmentation results， reduce the sensitivity of single scale features to the segmentation results， improve the robustness and stability of the segmentation algorithm， enhance the segmentation boundary sharpness and improve the segmentation accuracy. Dual?path multi?information domain attention module is responsible for extracting multi?scale features of 3D visual images by combining frequency domain channel and spatial attention mechanism， so as to reduce the sensitivity of single scale features to segmentation results. A hollow convolution layer is added in the multi?scale feature fusion module to increase the receptive field of multi?scale features， and the multi?scale features of the enlarged receptive field are fused to capture the detailed information and global information of the image， reduce the impact of noise on the segmentation results， and improve the robustness and stability of the segmentation algorithm. The Softmax classifier is used to process the fused features to obtain the scene segmentation results of 3D visual images. By fully?connected conditional random fields （CRFs）， the segmentation results are post?processed to optimize the segmentation boundary sharpness and improve the segmentation accuracy. Experimental results show that the proposed algorithm can effectively extract multi?scale features of 3D visual images and segment the scene of 3D visual images， and the boundary of scene segmentation is very clear. It is a new idea for the processing and analysis of 3D visual images.

Keywords： multi?scale; feature fusion; 3D vision; image scene segmentation; attention mechanism; dilated convolution; Softmax classifier; CRF

0 "引 "言

三維視覺圖像處理與分析技術(shù)不僅推動無人駕駛、智慧城市、虛擬現(xiàn)實、增強現(xiàn)實等新興技術(shù)的快速發(fā)展[1]，還拓寬了計算機視覺技術(shù)在工業(yè)檢測、醫(yī)療影像、安防監(jiān)控等各個領(lǐng)域的應用范圍[2]。三維視覺圖像場景分割作為計算機視覺領(lǐng)域中的一項關(guān)鍵技術(shù)，旨在將三維場景圖像劃分為多個具有相似性質(zhì)的區(qū)域或?qū)ο?，實現(xiàn)對場景內(nèi)容的精細理解和分析[3]?！?br>