当前位置: X-MOL 学术Int. J. Inf. Technol. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Hybrid descriptor definition for content based image classification using fusion of handcrafted features to convolutional neural network features
International Journal of Information Technology Pub Date : 2021-06-22 , DOI: 10.1007/s41870-021-00722-x
Rik Das , Khusbu Kumari , Sourav De , P. K. Manjhi , Sudeep Thepade

The efficacy of content-based image classification is dependent on the richness of the feature vectors extracted from the image data. Traditional feature extraction techniques highlight single low level image characteristics like shape, size, texture, color etc. for feature generation. The process often fails to extract meaningful descriptors since considering a single image characteristic will ignore other rich properties of image contents. Mass adoption of Convolutional Neural Network (CNN) has significantly improved classification performances for content-based image data. Recent literature has documented high level of precision for image classification by carrying out transfer learning with open source pre trained Convolutional Neural Network (CNN). However, the concept of transfer learning experiences major setback in case of limited training data due to overfitting of training instances during finetuning. This work has identified this challenge and has attempted to capture probability distribution of input images to a pre trained Convolution Neural Network by utilizing it as a fixed weight feature extractor. The proposed approach has addressed the overfitting problem by removing the finetuning step involved in transfer learning. Further investigation on robust descriptor definition is carried out by concatenating the pre-trained CNN features to handcrafted features. The hybrid architecture has encouraging outcomes that have outclassed the classification accuracies of state-of-the-art handcrafted techniques.



中文翻译:

使用手工特征与卷积神经网络特征融合的基于内容的图像分类的混合描述符定义

基于内容的图像分类的有效性取决于从图像数据中提取的特征向量的丰富程度。传统的特征提取技术突出单个低级图像特征,如形状、大小、纹理、颜色等,用于特征生成。该过程通常无法提取有意义的描述符,因为考虑单个图像特征将忽略图像内容的其他丰富属性。卷积神经网络 (CNN) 的大规模采用显着提高了基于内容的图像数据的分类性能。最近的文献记录了通过使用开源预训练卷积神经网络 (CNN) 进行迁移学习来实现图像分类的高精度。然而,由于微调期间训练实例过度拟合,在训练数据有限的情况下,迁移学习的概念会遇到重大挫折。这项工作已经确定了这一挑战,并尝试将输入图像的概率分布捕获到预训练的卷积神经网络,将其用作固定权重特征提取器。所提出的方法通过去除迁移学习中涉及的微调步骤解决了过拟合问题。通过将预训练的 CNN 特征连接到手工制作的特征,对稳健的描述符定义进行了进一步研究。混合架构取得了令人鼓舞的结果,其分类准确度超过了最先进的手工技术。这项工作已经确定了这一挑战,并尝试将输入图像的概率分布捕获到预训练的卷积神经网络,将其用作固定权重特征提取器。所提出的方法通过去除迁移学习中涉及的微调步骤解决了过拟合问题。通过将预训练的 CNN 特征连接到手工制作的特征,对稳健的描述符定义进行了进一步研究。混合架构取得了令人鼓舞的结果,其分类准确度超过了最先进的手工技术。这项工作已经确定了这一挑战,并尝试将输入图像的概率分布捕获到预训练的卷积神经网络,将其用作固定权重特征提取器。所提出的方法通过去除迁移学习中涉及的微调步骤解决了过拟合问题。通过将预训练的 CNN 特征连接到手工制作的特征,对稳健的描述符定义进行了进一步研究。混合架构取得了令人鼓舞的结果,其分类准确度超过了最先进的手工技术。所提出的方法通过去除迁移学习中涉及的微调步骤解决了过拟合问题。通过将预训练的 CNN 特征连接到手工制作的特征,对稳健的描述符定义进行了进一步研究。混合架构取得了令人鼓舞的结果,其分类准确度超过了最先进的手工技术。所提出的方法通过去除迁移学习中涉及的微调步骤解决了过拟合问题。通过将预训练的 CNN 特征连接到手工制作的特征,对稳健的描述符定义进行了进一步研究。混合架构取得了令人鼓舞的结果,其分类准确度超过了最先进的手工技术。

更新日期:2021-06-22
down
wechat
bug