摘要
随着电子商务的迅猛发展，网络购物已经成为人们日常生活中不可或缺的一部分。但随着在线商品种类的逐渐增多和商品图像的海量增长，基于文本特征的商品检索系统由于文本标注的局限性，在查准率方面很难满足用户的需求。而基于图像特征的检索方式利用图像自身特征来表示图像内容，弥补了文本检索的缺点，提高了商品分类和检索的精确度。但是在实际生活中，商品图像属性众多、背景复杂，同时容易受到遮挡、光照等因素的影响，使用传统的图像特征提取方法并不能排除这些因素的干扰，因此如何对商品进行准确地描述并以此提高商品检测精度是商品检索领域的研究热点。
近年来，卷积神经网络在商品检索领域的应用和发展受到了许多学者的关注，并且取得了很大进展，但仍存在一些问题。因此本文针对这些问题提出基于卷积神经网络的多尺度商品图像目标检测算法，主要研究内容如下：
1、由于商品图像背景十分复杂，会对商品的检测识别造成干扰，因此对目标区域的准确提取是商品检测识别的重要前提。本文首先采用全卷积网络提取人体区域，减少复杂背景对服装分割的影响。然后在分割出的人体区域上利用多目标检测框架提取服装区域和服装上的图案区域，为后续的特征提取和商品的分类识别奠定良好的基础。
2、在对商品图像进行分割的基础上，为提高对商品多属性特征的学习能力，将多任务学习的方法引入到
网络结构中。此外，由于线下商品图像受光照等因素的影响较大，引入度量学习中的Triplet结构，通过线上线下服装对的学习，提高网络对线下服装商品的识别精度。
3、为提高多尺度商品图像检测与识别的精确度，本文同时对服装商品区域和服装上的图案区域进行检测和识别，利用残差网络从多个尺度对商品特征进行描述，以便于能够更准确的对目标商品进行识别分类并有效地解决因尺度大小造成的商品区域不完整等问题。
本文针对现有商品目标检测中存在的问题分别从图像预处理、目标区域提取和特征识别三方面进行改进。最后通过实验从不同角度与其他算法进行对比分析，相比主流算法能够有效提高准确率，验证了本文算法在商品目标检测方面的优越性。
关键词：卷积神经网络，多尺度，目标检测，图像分割
Multi - Scale Commodity Targets Detection Algorithm Bad on考研辅导班收费
Convolutional Neural Network
Zhou Jinghong（Computer Science and Technology）
Directed by Associate Prof. Gong An
红裹肚Abstract
大学生创业贷款
With the rapid development of E-Commerce, online shopping has become an indispensable part of people's daily life. However, with the increasing number of online merchandi types and the massive increa of merchandi images, the text-bad product arch system is difficult to satisfy the ur's needs in terms of accuracy becau of the limitation of text labeling. The retrieval method bad on image feature us the features of the image itlf to express the image content, which makes up for the shortcomings of text retrieval and improves the accuracy of product classification and retrieval. However, in real life, the product image has many attributes and the background is complex. At the same time, it is easily affected by factors such as occlusion and illumination. Using traditional image feature extraction methods cannot eliminate the interference of the factors. Therefore, how to accurately describe the product and u it to increa product detection accuracy is a rearch hotspot in the field of product arch.
In recent years, the application and development of convolutional neural networks in the field of commodity retrieval has attracted the attention of many scholars, and has made great progress, but there are still some problems. Therefore, this paper propos a multi-scale commodity image target detection algorithm bad on convolutional neural network for the problems. The main rearch c和答元明黔南赠别
ontents are as follows:
1. Becau the background of the product image is very complex, it will interfere with the detection and recognition of the product. Therefore, accurate extraction of the target area is an important prerequisite for product detection and recognition. Firstly, this paper us the full convolutional network to extract the human body area, reducing the impact of complex background on clothing gmentation. Then the multi-target detection frame is ud to extract
the pattern area on the clothing area and clothing on the gmented human body area, which lays a good foundation for subquent feature extraction and product classification and recognition.一只船
2. Bad on the gmentation of product images, in order to improve the learning ability of multi-attribute features of products, the multi-task learning method is introduced into the network structure. In addition, becau offline merchandi images are greatly affected by light and other factors, the triplet structure in metric learning is introduced. Through the online and offline clothing pair learning, the recognition accuracy of online clothing goods under the line is improved.
3. To improve the accuracy of multi-scale product image detection and recognition, this article simultaneously detects and recognizes apparel product areas and clothing pattern areas, and us t
he residual network to describe product features from multiple scales in order to be able to Accurately identify and classify the target products and effectively solve problems such as incomplete product areas due to size.
In this paper, the existing problems in the detection of existing goods targets are improved from three aspects: image preprocessing, target area extraction and feature recognition. Finally, through experiments comparing with other algorithms from different angles, compared with mainstream algorithms can effectively improve the accuracy rate, verifying the superiority of the algorithm in commodity target detection.
Key words：convolutional neural network, multi-scale, target detection, image gmentation
目录
第一章绪论 (1)
1.1 研究背景及意义 (1)
1.2 国内外研究现状 (2)
爱情诗句1.2.1 基于传统图像处理的方法 (3)
1.2.2 基于卷积神经网络的方法 (4)
1.3 主要研究内容 (5)
1.4 章节安排 (6)
第二章相关知识和技术基础 (7)
2.1 卷积神经网络理论基础 (7)
2.1.1 卷积层 (8)
2.1.2 池化层 (9)
2.1.3 动机 (10)
2.2 物体检测主要方法 (11)
2.2.1 基于图像分割的物体检测方法 (12)
2.2.2 基于机器学习的物体检测方法 (14)
2.2.3 基于特征点的物体检测方法 (15)
2.2.4 基于背景减除的物体检测方法 (16)
2.3 物体识别主要方法 (16)
2.3.1 基于几何特征的物体识别方法 (16)
2.3.2 基于神经网络的物体识别方法 (17)
2.3.3 基于线性子空间的物体识别方法 (18)
2.3.4 基于颜色特征的物体识别方法 (19)
2.4 本章小结 (19)
第三章多尺度服装商品图像检测 (21)
3.1 基于全卷积网络的图像语义分割 (21)
3.2 多尺度商品目标区域提取 (24)
3.3 本章小结 (30)
第四章商品目标图像特征提取与识别 (31)
4.1 图像特征提取和识别 (31)
4.2 多类别属性分类 (34)
4.2.1 多任务学习 (34)
拔模斜度4.2.2 采用多任务学习的ResNet (36)
4.3 度量学习 (37)
4.3.1 Triplet结构网络模型 (38)
4.3.2 结合度量学习的ResNet (39)
4.4 本章小结 (40)
第五章实验分析 (41)
5.1 服装商品图像数据集 (41)
5.2 迁移学习 (43)
5.3 服装商品图像检测结果 (44)
5.4 服装商品图像识别结果 (45)
5.5 本章小结 (48)
门禁安装总结与展望 (49)
参考文献 (51)
攻读硕士学位期间取得的学术成果 (58)
致谢 (59)
