Data Augmentation Based on Color Features for Limited Training Texture Classification

Duong, Huu-Thanh; Hoang, Vinh Truong

doi:10.1109/incit.2019.8911934

Cited by 6 publications

(8 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…To evaluate the proposed method, we use three texture image datasets such as Outex-13 [24], USP-tex [25], and Stex [15] in Figure 2. These datasets have the test suites for evaluating in supervised learning context.…”

Section: Resultsmentioning

confidence: 99%

“…To extract color LBP features, many works [21], [22] have been proposed to consider the interaction between pixel from different color components. In this work, we follow [15] to extract extend opponent color LBP (OCLBP), each image is then characterized by 9 LBP histograms.…”

Section: Local Binary Patternsmentioning

confidence: 99%

“…Instead of using any pre-processing methods such as translation, compression, flipping for creating synthesis images, Duong and Hoang [15] recently applied color space transformation for generating and extracting local binary pattern (LBP) color features. They only used one training image for each class and then generate new image by transforming that image into another color space.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Data augmentation by combining feature selection and color features for image classification

Meethongjan

Hoang

Surinwarangkoon

2022

IJECE

Self Cite

View full text Add to dashboard Cite

<span lang="EN-US">Image classification is an essential task in computer vision with various applications such as bio-medicine, industrial inspection. In some specific cases, a huge training data is required to have a better model. However, it is true that full label data is costly to obtain. Many basic pre-processing methods are applied for generating new images by translation, rotation, flipping, cropping, and adding noise. This could lead to degrade the performance. In this paper, we propose a method for data augmentation based on color features information combining with feature selection. This combination allows improving the classification accuracy. The proposed approach is evaluated on several texture datasets by using local binary patterns features.</span>

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Local Binary Patternsmentioning

confidence: 99%

mentioning

confidence: 99%

See 1 more Smart Citation

Data augmentation by combining feature selection and color features for image classification

Meethongjan

Hoang

Surinwarangkoon

2022

IJECE

Self Cite

View full text Add to dashboard Cite

show abstract

“…Phương pháp làm tăng thêm dữ liệu đầu vào cho mô hình là một trong những phương pháp ít tốn kém nhưng hiệu quả để giải quyết vấn đề này. Việc làm tăng thêm dữ liệu này được áp dụng rộng rãi trong các bài toán thị giác máy tính (Wang & Luis, 2017) bằng cách sử dụng những kỹ thuật đơn giản như lật hình, xoay hình, cắt hình, thay đổi tỷ lệ ảnh hoặc biến đổi màu sắc (Duong & Truong, 2019b) nhằm thay đổi hình ảnh ban đầu. Do sự phức tạp về mặt ngữ nghĩa, sự đa dạng về mặt ngữ pháp và ngữ cảnh của ngôn ngữ, cho nên phương pháp làm tăng thêm dữ liệu đối với bài toán sử dụng dữ liệu văn bản vẫn còn là vấn đề nhiều thách thức.…”

Section: Giới Thiệuunclassified

Kỹ thuật làm tăng dữ liệu trong phân tích cảm xúc trên ngôn ngữ tiếng Việt

Thien¹

2022

HCMCOUJS - KỸ THUẬT VÀ CÔNG NGHỆ

View full text Add to dashboard Cite

Những bình luận phản hồi trong các hệ thống trực tuyến là một nguồn dữ liệu mang nhiều thông tin, cảm xúc của khách hàng về những sản phẩm hoặc dịch vụ. Những thông tin này được khai thác nhằm đem lại những ích lợi trong việc hoạch định chiến lược, quản trị khách hàng. Để đạt được những kết quả tốt đối với mô hình phân tích cảm xúc, đòi hỏi một lượng lớn dữ liệu được gán nhãn. Chi phí cho việc gán nhãn dữ liệu huấn luyện bởi con người là rất lớn. Trong nghiên cứu này chúng tôi đề xuất một mô hình làm tăng dữ liệu văn bản dựa trên các câu bình luận áp dụng cho ngôn ngữ tiếng Việt. Một số kỹ thuật cơ bản được sử dụng nhằm sinh thêm số lượng bình luận như chèn từ, thay thế từ, xóa từ. Kết quả thực nghiệm đã cho thấy hiệu quả của mô hình này.

show abstract

“…This kind of augmentation is helpful for making models capture similar signals, but from different viewpoints. Other types of augmentation will focus on changing color characteristics such as lighting, contrast, hue, and saturation [29]. This class of augmentation is helpful to make models more color agnostic and focus on shape features.…”

Section: Data Augmentationmentioning

confidence: 99%

Augmenting biological pathway extraction with synthetic data and active learning

Thompson¹

View full text Add to dashboard Cite

The corpus of biomedical literature is growing rapidly as many papers are recorded in PubMed every day. These papers often contain high-quality biological pathways in their figures/text, which are great resources for studying biological mechanisms and precision medicine. However, it can take a long time for many of these works to be put into practical use as each paper's contributions need to be curated by experts. This, often lengthy, process causes professional practice to lag behind research. To speed up this process, I helped develop a pipeline that integrates NLP and object detection processing to extract gene relationships reported in articles' figures and text. This pipeline was able to extract such relationships with high precision and recall on a small, annotated set. However, extending this pipeline for improved generalization and new settings was limited by the number of high-quality annotations available. Such labeled data is very time consuming to collect and traditional augmentations were observed to generate diminishing returns. To address this shortcoming, I developed an approach for generating purely synthetic data for object detection on biological pathway diagrams based on a set of rules and domain knowledge. Our method iteratively generates each pathway relationship uniquely and is demonstrated to improve the generalization of our object detection model significantly across a variety of settings. Additionally, with the capability to generate unique and informative samples, we integrated our synthetic generation methodology into an active learning setting. While traditional active learning relies on a pool of unlabeled data to draw from with an acquisition function, our method is pool-less and does not require any acquisition function. Instead, we generate each batch of data uniquely based on the training losses from the previous batch. Pool-less Active Learning (PAL) via synthetic data generation is demonstrated to reduce the number of iterations required for model convergence during training on pathway figures.

show abstract

Data Augmentation Based on Color Features for Limited Training Texture Classification

Cited by 6 publications

References 17 publications

Data augmentation by combining feature selection and color features for image classification

Data augmentation by combining feature selection and color features for image classification

Kỹ thuật làm tăng dữ liệu trong phân tích cảm xúc trên ngôn ngữ tiếng Việt

Augmenting biological pathway extraction with synthetic data and active learning

Contact Info

Product

Resources

About