融合可变形卷积网络的细粒度图像识别研究
2021-08-03吴忠粱
吴忠粱



摘要:针对细粒度图像识别领域中识别率不高、难以定位到图像中具有表征性的局部区域的问题,提出一种基于多区域融合的可变形卷积网络算法,该算法采用新型的卷积计算方式,根据越靠近边缘的部位越发包含更多图像上下文信息的原理,对图像给定多个中心并划分权重区域。在主流数据集上的实验结果表明,提出的基于多区域的可变形卷积网络结构在细粒度图像识别上的表现相比其他主流算法都有了一定的提升,并且相比于原始的可变形卷积网络和v2版本的可变形卷积模型也有了性能上的优化。
关键词:细粒度图像识别;局部表征性;多区域可变形卷积网络;图像上下文信息;区域划分
中图分类号:TP301 文献标识码:A
文章编号:1009-3044(2021)17-0193-03
开放科学(资源服务)标识码(OSID):
Fine-grained Image Classification Research Fused with Deformable Convolutional Network
WU Zhong-liang
(School of Information Engineering, East China University of Technology, Nanchang 30013,China)
Abstract: Aiming at the problem that the recognition rate is not high in the field of fine-grained image recognition and it is difficult to locate the characteristic local area in the image, a multi-region-based deformable convolution network algorithm is proposed, which uses a new convolution calculation method. According to the principle that the parts closer to the edge contain more image context information, multiple centers are given to the image and weighted regions are divided. Experimental results on mainstream data sets show that the performance of the proposed multi-region-based deformable convolutional network structure in fine-grained image recognition has a certain improvement compared with other mainstream algorithms, and compared to the original deformable The convolutional network and the v2 version of the deformable convolution model have also been optimized in performance.
Key words: fine-grained image recognition; local representation; multi-region deformable convolutional network; image context information; region partition
1 引言
在计算机视觉领域中,图像分类任务通常分为粗粒度图像分类和细粒度图像分类,其中细粒度图像分类要求识别出某一基类下的数百种子类,包括识别不同种类的鸟[1],车[2],宠物[3],花[4],飞行器[5]等,而粗粒度图像分类只需要识别出图像中目标的基本类别。现如今细粒度图像分类这一任务具有非常迫切的研究需求,并且其應用非常广泛,例如生态多样性的保护、自动驾驶以及癌症检测等。
相比一般的图像识别,细粒度图像识别难度较大。一方面,由于部分目标在外观上只存在细微的差异,例如鸟类在形状,背部颜色和纹理特征上的细微差异,而这种细微的局部的差异恰恰是细粒度图像分类的关键所在。……
