这个作业是完成图像识别在食品中使用研究报告的paper代写

 

ENGR 295A Proposal Template

1问题或机会的识别和意义(简)

1.1机器学习在食品图像识别中的意义

食品图像识别无论在实践上还是技术上都至关重要。食物和饮食是我们日常生活中最受欢迎的主题之一。有许多与食物有关的APP,例如为食物拍照,帮助保持饮食习惯和计算卡路里。如果有有效的识别方法,这些应用都会受益匪浅。
从技术的角度来看,食物图像识别比许多其他众所周知的识别问题(例如人脸识别)要困难得多,因为其形状和颜色各异且带有嘈杂的背景。因此,如果我们能够解决领域对象识别中的一些常见问题,那么从事食品图像识别的工作也会从技术上使许多其他领域受益。

1.2机器学习在食品图像识别中的机会

除了挑战之外,该领域还有巨大的潜力。公司正在研究可以在移动APP中实现的更好的算法,这导致该技术的高商业价值。食物识别的另一个优点是,与医学研究等某些领域相比,食物的图像数据集更容易构建,这使得在输出方面取得进展变得更加容易。

2确定假设(简)

2.1结合不同的机器学习技术将进一步提高检测精度

从以前的工作中,我们发现食品图像识别中使用的机器学习技术已经发展了多年。在使用深度学习技术之前,人们使用诸如支持向量机之类的传统方法。当卷积神经网络于2014年问世时,它在图像处理项目中显示出巨大的潜力,Kagaya [1]首先将CNN纳入了食品图像识别,其结果超过了以前的方法,将准确率从60%提高到90%以上。
在机器学习项目中,最具挑战性的部分之一是输入训练数据的质量。近年来,在图像处理领域引入了新的与深度学习相关的方法,例如图像分割,图像数据增强,并被证明可以有效地提高训练效果。数据准备中图像分割的基本思想是从背景中提取某些对象,以减少训练时的噪声数据。当输入数据集的大小不够大而无法进行越来越复杂的训练过程时,数据增强将提供帮助。在食物图像识别中,由于食物图像始终被杂乱的炊具包围或目标对象被包围,人们也遭受上述问题的困扰有时大小不足以提供有意义的像素。我们假设通过在输入数据集上包括新开发的图像预处理方法,我们将获得比以前的CNN模型更好的结果。

2.2 Improved training methods would strengthen the ability of Food Image Detection on multiple types of food

Training a large Convolutional Neural Network has always been a big issue which stops people from applying the powerful deep learning techniques in more fields. To solve the problem of computational intenseness, many different transfer learning methods have been involved. Based on open-source large-scale image datasets, Maxime Oquab[2] developed methods to build CNN on specific domains without large quantities of images available.

Another advantage of transfer learning is that recent work of Kabir[3] shows that machine learning algorithms can benefit from data of different domains. Based on the discovery, it’s worth testing tuning some well-trained CNN models on food image datasets, even though the underlying data is of other specific fields such as human faces.

Due to the lack of high-quality food images and the limitation of computation power, Kagaya’s CNN model can only recognize 10 types of food, which is far from enough if we are going to put the model into practical use such as being implemented in diet APPs. We assume that by introducing transfer learning and well-trained CNN models from other domains, our new model could outweigh the former one.

4 Feasibility Study/Analysis (Kate)
same as we did in ENGR 202 system engineering
4.1 Subsection
4.2 Subsection
5 Project Concepts and High Level Architectures (Samantha)
block diagrams, high level architecture (your own – CNN, volume estimator, calorie calculator)

4.1 Project Concepts. We accomplish system planning and architecting in response to the customer needs. Our project requirements are to be able to detect fast food images, to accurately measure daily food intake, and to provide a detailed food profile with nutrition and calories information. Our project’s functional requirements have a three-step process, such as extracting regions of interests (ROIs), classifying different food categories, and generating a dietary assessment report. System Operational Requirements are for the first-step process, apply Region Proposal Network (RPN) which is derived from Faster R-CNN model to separate food items from the image background, for the second-step process, apply deep Convolutional Neural Network (CNN) on selected ROIs for image processing, segmentation, and classification into different food categories, and for the third-step process, use dietary assessment tools to provide nutrition analysis report. For Functional Architecture, we will ensure that our application will satisfy the software requirements’ specifications and stakeholders’ needs and expectations.


EasyDue™ 支持PayPal, AliPay, WechatPay, Taobao等各种付款方式!

E-mail: easydue@outlook.com  微信:easydue


EasyDue™是一个服务全球中国留学生的专业代写公司
专注提供稳定可靠的北美、澳洲、英国代写服务
专注提供CS、统计、金融、经济、数学等覆盖100+专业的作业代写服务