Object Classifications by Image Super-Resolution Preprocessing for Convolutional Neural Networks

Na, Bokyoon; Fox, Geoffrey

doi:10.25046/aj050261

Cited by 11 publications

(7 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To efficiently train and test CNNs, a series of pre-processing procedures normally need to be applied (Na & Fox, 2020). Specifically, before model training all images were resized to the same resolution (473 x 372 pixels) by taking into account the mean resolution of the set, and then normalized to the [0,1] range.…”

Section: Model Trainingmentioning

confidence: 99%

Deep learning assessment of cultural ecosystem services from social media images

Renna

Moreno-Llorca

et al. 2021

Preprint

View full text Add to dashboard Cite

Crowdsourced social media data has become popular in the assessment of cultural ecosystem services (CES). Advances in deep learning show great potential for the timely assessment of CES at large scales. Here, we describe a procedure for automating the assessment of image elements pertaining to CES from social media. We focus on a binary (natural, human) and a multiclass (posing, species, nature, landscape, human activities, human structures) classification of those elements using two Convolutional Neural Networks (CNNs; VGG16 and ResNet152) with the weights from two large datasets - Places365 and ImageNet -, and our own dataset. We train those CNNs over Flickr and Wikiloc images from the Peneda-Geres region (Portugal) and evaluate their transferability to wider areas, using Sierra Nevada (Spain) as test. CNNs trained for Peneda-Geres performed well, with results for the binary classification (F1-score > 80%) exceeding those for the multiclass classification (> 60%). CNNs pre-trained with Places365 and ImageNet data performed significantly better than with our data. Model performance decreased when transferred to Sierra Nevada, but their performances were satisfactory (> 60%). The combination of manual annotations, freely available CNNs and pre-trained local datasets thereby show great relevance to support automated CES assessments from social media.

show abstract

Section: Model Trainingmentioning

confidence: 99%

Deep learning assessment of cultural ecosystem services from social media images

Renna

Moreno-Llorca

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…The CNNs considered as the magic solution for much computer vision problems. The CNNs depend on some of filters that reduce the image height and width and increase the number of channels, then processing the output with full connected (FC) neural network layers [22][23][24]. Figure 1 shows one of the CNNs model structures [23].…”

Section: The Convolution Neural Network (Cnns)mentioning

confidence: 99%

Investment of Classic Deep CNNs and SVM for Classifying Remote Sensing Images

Al-Afandy¹,

Johra²,

Lazaar³

et al. 2020

Adv. sci. technol. eng. syst. j.

View full text Add to dashboard Cite

Feature extraction is an important process in image classification for achieving an efficient accuracy for the classification learning models. One of these methods is using the convolution neural networks. The use of the trained classic deep convolution neural networks as features extraction gives a considerable results in the remote sensing images classification models. So, this paper proposes three classification approaches using the support vector machine where based on the use of the ImageNet pre-trained weights classic deep convolution neural networks as features extraction from the remote sensing images. There are three convolution models that used in this paper; the Densenet 169, the VGG 16, and the ResNet 50 models. A comparative study is done by extract features using the outputs of the mentioned ImageNet pre-trained weights convolution models after transfer learning, and then use these extracted features as input features for the support vector machine classifier. The used datasets in this paper are the UC Merced land use dataset and the SIRI-WHU dataset. The comparison is based on calculating the overall accuracy to assess the classification model performance.

show abstract

“…The CNNs are taken from the ANNs with exception that it is not fully connected layers. The CNNs are the best solution for computer vision which based on some of filters to reduce the image height and width and increase the number of channels together, then processing the output with full connected neural network layers (FCs) which reduce the input layer neurons, reduce training time, and increase the training model performance [24][25][26][27]. These filters values are initialized with many random functions which can be optimized.…”

Section: The Convolution Neural Network (Cnns)mentioning

confidence: 99%

“…These filters values are initialized with many random functions which can be optimized. The filters design is based on the use of self-intuition as with ANNs structure design and learning hyper-parameters coefficient choice, which increasing the difficulty to reach the best solution for the learning problems [26][27]. Figure 1 shows an example of a CNNs model structure [25].…”

Section: The Convolution Neural Network (Cnns)mentioning

confidence: 99%

Using Classic Networks for Classifying Remote Sensing Images: Comparative Study

Al-Afandy¹,

Lazaar²,

Achhab³

2020

Adv. sci. technol. eng. syst. j.

View full text Add to dashboard Cite

This paper presents a comparative study for using the deep classic convolution networks in remote sensing images classification. There are four deep convolution models that used in this comparative study; the DenseNet 196, the NASNet Mobile, the VGG 16, and the ResNet 50 models. These learning convolution models are based on the use of the ImageNet pretrained weights, transfer learning, and then adding a full connected layer that compatible with the used dataset classes. There are two datasets are used in this comparison; the UC Merced land use dataset and the SIRI-WHU dataset. This comparison is based on the inspection of the learning curves to determine how well the training model is and calculating the overall accuracy that determines the model performance. This comparison illustrates that the use of the ResNet 50 model has the highest overall accuracy and the use of the NASNet Mobile model has the lowest overall accuracy in this study. The DenseNet 169 model has little higher overall accuracy than the VGG 16 model.

show abstract

Object Classifications by Image Super-Resolution Preprocessing for Convolutional Neural Networks

Cited by 11 publications

References 14 publications

Deep learning assessment of cultural ecosystem services from social media images

Deep learning assessment of cultural ecosystem services from social media images

Investment of Classic Deep CNNs and SVM for Classifying Remote Sensing Images

Using Classic Networks for Classifying Remote Sensing Images: Comparative Study

Contact Info

Product

Resources

About