基于多尺度特征融合的彩色图像检索算法  

Color Image Retrieval Algorithm Based on Multi-scale Feature Fusion

作　　者：苟权海 韩成功 王伟臣 寇旗旗 江曼 GOU Quanhai;HAN Chenggong;WANG Weichen;KOU Qiqi;JIANG Man(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou Jiangsu 221116,China;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou Jiangsu 221116,China)

机构地区：[1]中国矿业大学信息与控制工程学院,江苏徐州221116 [2]中国矿业大学计算机科学与技术学院,江苏徐州221116

出　　处：《佳木斯大学学报（自然科学版）》2023年第3期1-6,共6页Journal of Jiamusi University：Natural Science Edition

基　　金：国家重点研发计划项目(2021YFC2902702)。

摘　　要：针对检索过程中场景复杂、目标具有相似性影响检索准确性的问题,提出了基于多尺度特征融合的彩色图像检索算法。首先,选用Inception-v4网络提取图像的多尺度深度特征,在此基础上引入通道注意力机制,通过自适应调整权重,改善对于各通道的特征提取效果;接着,引入空间注意力机制,在获取通道信息的基础上结合图像空间域的特征,定位目标位置,提升网络的检索性能;通过CP-FHOG方法提取图像的纹理信息和颜色信息,使用特征拼接层将图像的颜色纹理特征和深度特征进行有效融合,获取图像更加细致的全局特征,以抵抗数据集中的复杂场景和相似目标,提高检索的效果。在Corel-1000数据集上与先进的图像检索方法进行对比实验,实验结果证明,该算法相比其他算法准确率更高,尤其是在场景复杂、目标具有相似性的图像类别中,取得了更好的检索效果。Aiming to solve the problems of complex scene and similarity of objects affecting the retrieval accuracy,this paper proposes a color image retrieval algorithm based on multi-scale feature fusion.Firstly,the multi-scale depth features of images are extracted by using a Inception-v4 network.On this basis,channel attention mechanism is introduced to improve the feature extraction effect of each channel through adaptive weight adjustment.Secondly,the spatial attention mechanism is introduced to locate the target location and improve the retrieval performance of the network based on the acquisition of channel information and the characteristics of image spatial domain.Finally,the feature splicing layer is used to effectively fuse the color,texture and depth features of the image to obtain more detailed global features of the image,so as to resist complex scenes and similar targets in the data set and improve the retrieval effect.Experiments on Corel-1000 data sets show that the proposed algorithm has higher accuracy than other algorithms,especially in image categories with complex scenes and similar objects,and achieves better retrieval results.

关 键 词：图像检索 多尺度 特征融合 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]