A. D. Freitas, Adriano B. Silva, A. S. Martins, L. A. Neves, T. A. A. Tosta, P. D. Faria, M. Z. Nascimento
Cancer is one of the diseases with the highest mortality rate in the world. Dysplasia is a difficult-to-diagnose precancerous lesion, which may not have a good Hematoxylin and Eosin (H&E) stain ratio, making it difficult for the histology specialist to diagnose. In this work, a method for normalizing H&E stains in histological images was investigated. This method uses a generative neural network based on a U-net for image generation and a PatchGAN architecture for information discrimination. Then, the normalized histological images were employed in classification algorithms to investigate the detection of the level of dysplasia present in the histological tissue of the oral cavity. The CNN models as well as hybrid models based on learning features and machine learning algorithms were evaluated. The employment of the ResNet-50 architecture and the Random Forest algorithm provided results with an accuracy rate around 97% for the images normalized with the investigated method.
{"title":"Evaluation of normalization technique on classification with deep learning features","authors":"A. D. Freitas, Adriano B. Silva, A. S. Martins, L. A. Neves, T. A. A. Tosta, P. D. Faria, M. Z. Nascimento","doi":"10.5753/wvc.2021.18898","DOIUrl":"https://doi.org/10.5753/wvc.2021.18898","url":null,"abstract":"Cancer is one of the diseases with the highest mortality rate in the world. Dysplasia is a difficult-to-diagnose precancerous lesion, which may not have a good Hematoxylin and Eosin (H&E) stain ratio, making it difficult for the histology specialist to diagnose. In this work, a method for normalizing H&E stains in histological images was investigated. This method uses a generative neural network based on a U-net for image generation and a PatchGAN architecture for information discrimination. Then, the normalized histological images were employed in classification algorithms to investigate the detection of the level of dysplasia present in the histological tissue of the oral cavity. The CNN models as well as hybrid models based on learning features and machine learning algorithms were evaluated. The employment of the ResNet-50 architecture and the Random Forest algorithm provided results with an accuracy rate around 97% for the images normalized with the investigated method.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127525728","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
André Machado, K. Aires, R. Veras, L. B. Britto Neto
This paper proposes a new approach in object recognition to assist visually impaired people. This approach achieved accuracy rates higher than the approaches proposed by the authors of the selected datasets. We applied Data Augmentation with other techniques and adjustments to different Pre-trained CNNs (Convolutional Neural Networks). The ResNet-50 based approach achieved the best results in the most recent datasets. This work focused on products that are usually found on grocery store shelves, supermarkets, refrigerators or pantries.
{"title":"Grocery Product Recognition to Aid Visually Impaired People","authors":"André Machado, K. Aires, R. Veras, L. B. Britto Neto","doi":"10.5753/wvc.2021.18896","DOIUrl":"https://doi.org/10.5753/wvc.2021.18896","url":null,"abstract":"This paper proposes a new approach in object recognition to assist visually impaired people. This approach achieved accuracy rates higher than the approaches proposed by the authors of the selected datasets. We applied Data Augmentation with other techniques and adjustments to different Pre-trained CNNs (Convolutional Neural Networks). The ResNet-50 based approach achieved the best results in the most recent datasets. This work focused on products that are usually found on grocery store shelves, supermarkets, refrigerators or pantries.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"66 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115580861","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The outputs of CNN layers, called Activations, are composed of Feature Maps, which show textural information that can be extracted by a texture descriptor. Standard CNN feature extraction use Activations as feature vectors for object recognition. The goal of this work is to evaluate a new methodology of CNN feature extraction. In this paper, instead of using the Activations as a feature vector, we use a CNN as a feature extractor, and then we apply a texture descriptor directly on the Feature Maps. Thus, we use the extracted features obtained by the texture descriptor as a feature vector for authentication. To evaluate our proposed method, we use the AlexNet CNN previously trained on the ImageNet database as a feature extractor; then we apply the uniform LBP (uLBP) descriptor on the Feature Maps for texture extraction. We tested our proposed method on the VISOB dataset composed of periocular images taken from 3 different smartphones under 3 different lighting conditions. Our results show that the use of a texture descriptor on CNN Feature Maps achieves better performance than computer vision handcrafted methods or even by standard CNN feature extraction.
{"title":"Periocular authentication in smartphones applying uLBP descriptor on CNN Feature Maps","authors":"William Barcellos, A. Gonzaga","doi":"10.5753/wvc.2021.18890","DOIUrl":"https://doi.org/10.5753/wvc.2021.18890","url":null,"abstract":"The outputs of CNN layers, called Activations, are composed of Feature Maps, which show textural information that can be extracted by a texture descriptor. Standard CNN feature extraction use Activations as feature vectors for object recognition. The goal of this work is to evaluate a new methodology of CNN feature extraction. In this paper, instead of using the Activations as a feature vector, we use a CNN as a feature extractor, and then we apply a texture descriptor directly on the Feature Maps. Thus, we use the extracted features obtained by the texture descriptor as a feature vector for authentication. To evaluate our proposed method, we use the AlexNet CNN previously trained on the ImageNet database as a feature extractor; then we apply the uniform LBP (uLBP) descriptor on the Feature Maps for texture extraction. We tested our proposed method on the VISOB dataset composed of periocular images taken from 3 different smartphones under 3 different lighting conditions. Our results show that the use of a texture descriptor on CNN Feature Maps achieves better performance than computer vision handcrafted methods or even by standard CNN feature extraction.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"52 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116446748","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Skin cancer is one of the most common types of cancer in Brazil and its incidence rate has increased in recent years. Melanoma cases are more aggressive compared to nonmelanoma skin cancer. Machine learning-based classification algorithms can help dermatologists to diagnose whether skin lesion is melanoma or non-melanoma cancer. We compared four convolutional neural networks architectures (ResNet-50, VGG16, Inception-v3, and DenseNet-121) using different training strategies and validation methods to classify seven classes of skin lesions. The experiments were executed using the HAM10000 dataset which contains 10,015 images of pigmented skin lesions. We considered the test accuracy to determine the best model for each strategy. DenseNet-121 was the best model when trained with fine-tuning and data augmentation, 90% (k-fold crossvalidation). Our results can help to improve the use of machine learning algorithms for classifying pigmented skin lesions.
{"title":"A comparative study of convolutional neural networks for classification of pigmented skin lesions","authors":"Natalia Camillo do Carmo, J. F. Mari","doi":"10.5753/wvc.2021.18909","DOIUrl":"https://doi.org/10.5753/wvc.2021.18909","url":null,"abstract":"Skin cancer is one of the most common types of cancer in Brazil and its incidence rate has increased in recent years. Melanoma cases are more aggressive compared to nonmelanoma skin cancer. Machine learning-based classification algorithms can help dermatologists to diagnose whether skin lesion is melanoma or non-melanoma cancer. We compared four convolutional neural networks architectures (ResNet-50, VGG16, Inception-v3, and DenseNet-121) using different training strategies and validation methods to classify seven classes of skin lesions. The experiments were executed using the HAM10000 dataset which contains 10,015 images of pigmented skin lesions. We considered the test accuracy to determine the best model for each strategy. DenseNet-121 was the best model when trained with fine-tuning and data augmentation, 90% (k-fold crossvalidation). Our results can help to improve the use of machine learning algorithms for classifying pigmented skin lesions.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129708846","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pedro Henrique Silva Domingues, Renan Martins Mendes da Silva, Ibrahim Jamil Orra, Matheus Elias Cruz, T. Heiderich, C. Thomaz
The daily life of preterm babies may be involved with long exposure to pain, causing problems in the development of the nervous system. In this context, an on-going area of research is the scientific development of image-based automatic pain detection systems based on several techniques, from anatomical measurements to artificial intelligence, they have generally two main issues: the categorization of the most relevant facial regions for identifying neonatal pain and the practical difficulty related to the presence of artifacts obstructing parts of the face. This paper proposes and implements an areas-of-interest automatic segmentation method that allows the creation of a novel dataset containing crops of neonatal faces relevant for pain classification, labelled by areas-of-interest and pain status. Moreover, we have also investigated the use of similarity matching techniques to compare each area-of-interest to the corresponding one extracted from a prototype face with no occlusion.t
{"title":"Neonatal Face Mosaic: An areas-of-interest segmentation method based on 2D face images","authors":"Pedro Henrique Silva Domingues, Renan Martins Mendes da Silva, Ibrahim Jamil Orra, Matheus Elias Cruz, T. Heiderich, C. Thomaz","doi":"10.5753/wvc.2021.18914","DOIUrl":"https://doi.org/10.5753/wvc.2021.18914","url":null,"abstract":"The daily life of preterm babies may be involved with long exposure to pain, causing problems in the development of the nervous system. In this context, an on-going area of research is the scientific development of image-based automatic pain detection systems based on several techniques, from anatomical measurements to artificial intelligence, they have generally two main issues: the categorization of the most relevant facial regions for identifying neonatal pain and the practical difficulty related to the presence of artifacts obstructing parts of the face. This paper proposes and implements an areas-of-interest automatic segmentation method that allows the creation of a novel dataset containing crops of neonatal faces relevant for pain classification, labelled by areas-of-interest and pain status. Moreover, we have also investigated the use of similarity matching techniques to compare each area-of-interest to the corresponding one extracted from a prototype face with no occlusion.t","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133107317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Felipe Silveira Brito Borges, Juliana Velasques Balta, Milad Roghanian, A. B. Gonçalves, Marco A. Alvarez, H. Pistori
Palynology can be applied to different areas, such as archeology and allergy, where it is constantly growing. However, no publication comparing human classifications with machine learning classifications at different optical scales has been found in the literature. An image dataset with 17 pollen species that occur in Brazil was created, and machine learning algorithms were used for their automatic classification and subsequent comparison with humans. The experiments presented here show how machine and human classification behave according to different optical image scales. Satisfactory results were achieved, with 98.88% average accuracy for the machine and 45.72% for human classification. The results impact a single scale pattern for capturing pollen grain images for both future computer vision experiments and for a faster advance in palynology science.
{"title":"The interference of optical zoom in human and machine classification of pollen grain images","authors":"Felipe Silveira Brito Borges, Juliana Velasques Balta, Milad Roghanian, A. B. Gonçalves, Marco A. Alvarez, H. Pistori","doi":"10.5753/wvc.2021.18897","DOIUrl":"https://doi.org/10.5753/wvc.2021.18897","url":null,"abstract":"Palynology can be applied to different areas, such as archeology and allergy, where it is constantly growing. However, no publication comparing human classifications with machine learning classifications at different optical scales has been found in the literature. An image dataset with 17 pollen species that occur in Brazil was created, and machine learning algorithms were used for their automatic classification and subsequent comparison with humans. The experiments presented here show how machine and human classification behave according to different optical image scales. Satisfactory results were achieved, with 98.88% average accuracy for the machine and 45.72% for human classification. The results impact a single scale pattern for capturing pollen grain images for both future computer vision experiments and for a faster advance in palynology science.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"37 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133230822","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Raido Lacorte Galina, Thadeu Pezzin Melo, K. S. Komati
Cracks on the concrete surface are symptoms and precursors of structural degradation and hence must be identified and remedied. However, locating cracks is a time-consuming task that requires specialized professionals and special equipment. The use of neural networks for automatic crack detection emerges to assist in this task. This work proposes one U-Net based neural network to perform crack segmentation, trained with the Crack500 and DeepCrack datasets, separately. The U-Net used has seven contraction and seven expansion layers, which differs from the original architecture of four layers of each part. The IoU results obtained by the model trained with Crack500 was 71.03%, and by the model trained with DeepCrack was 86.38%.
{"title":"Pavement Crack Segmentation using a U-Net based Neural Network","authors":"Raido Lacorte Galina, Thadeu Pezzin Melo, K. S. Komati","doi":"10.5753/wvc.2021.18893","DOIUrl":"https://doi.org/10.5753/wvc.2021.18893","url":null,"abstract":"Cracks on the concrete surface are symptoms and precursors of structural degradation and hence must be identified and remedied. However, locating cracks is a time-consuming task that requires specialized professionals and special equipment. The use of neural networks for automatic crack detection emerges to assist in this task. This work proposes one U-Net based neural network to perform crack segmentation, trained with the Crack500 and DeepCrack datasets, separately. The U-Net used has seven contraction and seven expansion layers, which differs from the original architecture of four layers of each part. The IoU results obtained by the model trained with Crack500 was 71.03%, and by the model trained with DeepCrack was 86.38%.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115765053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Gabriel Peixoto de Carvalho, André Luiz Brandão, F. Ferreira
Despite the recent advancements in deep learning, sign language recognition persists as a challenge in computer vision due to its complexity in shape and movement patterns. Current studies that address sign language recognition treat hand pose recognition as an image classification problem. Based on this approach, we introduce HandArch, a novel architecture for realtime hand pose recognition from video to accelerate the development of sign language recognition applications. Furthermore, we present Libras91, a novel dataset of Brazilian sign language (LIBRAS) hand configurations containing 91 classes and 108,896 samples. Experimental results show that our approach surpasses the accuracy of previous studies while working in real-time on video files. The recognition accuracy of our system is 99% for the novel dataset and over 95% for other hand pose datasets.
{"title":"HandArch: A deep learning architecture for LIBRAS hand configuration recognition","authors":"Gabriel Peixoto de Carvalho, André Luiz Brandão, F. Ferreira","doi":"10.5753/wvc.2021.18883","DOIUrl":"https://doi.org/10.5753/wvc.2021.18883","url":null,"abstract":"Despite the recent advancements in deep learning, sign language recognition persists as a challenge in computer vision due to its complexity in shape and movement patterns. Current studies that address sign language recognition treat hand pose recognition as an image classification problem. Based on this approach, we introduce HandArch, a novel architecture for realtime hand pose recognition from video to accelerate the development of sign language recognition applications. Furthermore, we present Libras91, a novel dataset of Brazilian sign language (LIBRAS) hand configurations containing 91 classes and 108,896 samples. Experimental results show that our approach surpasses the accuracy of previous studies while working in real-time on video files. The recognition accuracy of our system is 99% for the novel dataset and over 95% for other hand pose datasets.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"56 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122127075","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this work, we used the Deep Learning (DL) architecture named U-Net to segment images containing side view cattle. We evaluated the ability of the U-Net to segment images captured with different backgrounds and from the different breeds, both acquired by us and from the Internet. Since cattle images present a more constant background than other applications, we also evaluated the performance of the U-Net when we change the numbers of convolutional blocks and filters. Results show that U-Net can be used to segment cattle images using fewer blocks and filters than traditional U-Net and that the number of blocks is more important than the total number of filters used.
{"title":"Unsupervised Segmentation of Cattle Images Using Deep Learning","authors":"Vinícius Guardieiro Sousa, A. Backes","doi":"10.5753/wvc.2021.18886","DOIUrl":"https://doi.org/10.5753/wvc.2021.18886","url":null,"abstract":"In this work, we used the Deep Learning (DL) architecture named U-Net to segment images containing side view cattle. We evaluated the ability of the U-Net to segment images captured with different backgrounds and from the different breeds, both acquired by us and from the Internet. Since cattle images present a more constant background than other applications, we also evaluated the performance of the U-Net when we change the numbers of convolutional blocks and filters. Results show that U-Net can be used to segment cattle images using fewer blocks and filters than traditional U-Net and that the number of blocks is more important than the total number of filters used.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"25 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129954111","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The coffee capsules brought practicality and speed in the preparation of the drink. However, with its popularization came a major environmental problem, the generation of a large amount of garbage, which for 2021 has an estimated 14 thousand tons of garbage, only coming from the capsules. To avoid this disposal it is necessary to recycle them, however it is not a trivial job, since they are composed of various materials, as well as the collection of these capsules presents challenges. Therefore, a collection system is of great value, which, in addition to being automated, generates bonuses proportional to the quantity of discarded capsules. This work is dedicated preliminary tests on the development of such a system using a convolutional neural network for the detection of coffee capsules. This algorithm was trained with two image sets, one containing images with reflection and the other without, which presented an accuracy of approximately 97%.
{"title":"Application of Convolutional Neural Network in Coffee Capsule Count Aiming Collection System for Recycling","authors":"Henrique Wippel Parucker da Silva, G. B. Santos","doi":"10.5753/wvc.2021.18907","DOIUrl":"https://doi.org/10.5753/wvc.2021.18907","url":null,"abstract":"The coffee capsules brought practicality and speed in the preparation of the drink. However, with its popularization came a major environmental problem, the generation of a large amount of garbage, which for 2021 has an estimated 14 thousand tons of garbage, only coming from the capsules. To avoid this disposal it is necessary to recycle them, however it is not a trivial job, since they are composed of various materials, as well as the collection of these capsules presents challenges. Therefore, a collection system is of great value, which, in addition to being automated, generates bonuses proportional to the quantity of discarded capsules. This work is dedicated preliminary tests on the development of such a system using a convolutional neural network for the detection of coffee capsules. This algorithm was trained with two image sets, one containing images with reflection and the other without, which presented an accuracy of approximately 97%.","PeriodicalId":311431,"journal":{"name":"Anais do XVII Workshop de Visão Computacional (WVC 2021)","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128821716","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：