Transfer learning application in a computer vision system for detection of driver distractionSouza, Bruno J.; Sobrinho, Sandro J. M.; Mayer, Fernando R.; Freire, Roberto Z.; Szejka, Anderson L.
doi: 10.1117/12.3031961pmid: N/A
The number of traffic accidents increases every year, and most of these accidents are caused by driver distraction. In countries with less developed road infrastructure, the number of accidents is considerably higher, just like in Brazil. Since distraction is one of the leading causes of accidents, there is a need for mechanisms that prevent drivers from becoming distracted. This paper shows the development of an intelligent image-based driver distraction detection system. Assuming interesting approaches considering neural networks (ANN) to solve the problem based on databases such as State Farm Distracted Driver Detection (SFD3) or AUC Distracted Driver V2 (AUCD2), this study aims to apply the transfer learning technique to obtain better performance and accuracy considering a smaller database. Assuming that the model must have a reduced architecture to be used in an embedded system, models based on convolutional neural networks (CNN) were chosen. Using transfer learning, it was possible to obtain a hit rate of 92.20% in AUCD2 and 64.47% considering the dataset proposed in this study.
EDSM: an encoder-decoder architecture face restoration network with style modulationTu, Yuchen; Jiang, Mengyao; Yu, Li
doi: 10.1117/12.3030002pmid: N/A
In recent years, face restoration methods based on deep learning with or without GAN prior have two main problems: retaining less identity information of the original input image and insufficient utilization of facial structure information. In order to solve the mentioned problems, we propose an encoder-decoder architecture face restoration network with style modulation called EDSM. First, skip connection and channel attention module are added to the basic network and a lightweight style modulation module is introduced to make full use of the global and local information extracted from the low-resolution (LR) face image. Meanwhile, identity loss is introduced to preserve identity information and a multi-scale discriminator is added to constitute the EDSM-plus network. Experiments have shown that the proposed EDSM and EDSM-plus have good face restoration performance in the Helen dataset.
Effects of hyper-parameters in online constrained clustering: a study on animal videosWilliams, Francis J.; Kuncheva, Ludmila I.
doi: 10.1117/12.3030009pmid: N/A
The aim of online clustering is to discover a structure in running data. Adding label constraints or pairwise constraints to this has shown to improve the clustering accuracy. In this study we present an analysis of how different hyperparameters – proportion of constraints, initial number of clusters, and batch window size – affect most recent and popular online constrained clustering methods, using three different metrics. Our results show that initial number of clusters and window size have an effect on clustering results, while the proportion of constraints does not. We also demonstrate that online clustering performs better than clustering of the whole data together. Our overall findings point at the need for new, more effective online constrained clustering methods.
Chinese multi-dialect speech recognition based on instruction tuningDing, Timin; Sun, Kai; Zhang, Xu; Yu, Jian; Huang, Degen
doi: 10.1117/12.3030013pmid: N/A
The technology of Chinese dialect speech recognition contributes to the preservation and inheritance of regional culture, as well as providing more convenient and customized services, with broad application prospects. In recent years, end-to-end speech recognition methods have demonstrated strong performance in dialect recognition. However, training the model using only a single dialect dataset would cause the model to lose the commonalities in acoustics and linguistics at a broader level. On the other hand, directly training a single model with multiple dialects would overlook the differences between dialect texts, thus affecting the model’s performance. To address this issue, this paper proposes a Chinese multi-dialect speech recognition method based on instruction tuning. By adding different instruction sets before different dialect texts, the model can learn the commonalities among different dialects within the same language while preserving the differences between dialect texts. Additionally, this paper also attempts to enhance the model’s text generation capability by using an additional language model for rescoring the model outputs. We conducted tests on the Common Voice dataset using the Whisper model. The results show that compared to the method of direct mixed training, the instruction finetuning method and rescoring method reduced the Word Error Rate (WER) by 13.44% and 21.18% respectively.
Extracting key drivers of sky ratings and evaluating air passenger's satisfaction classification model through online review analysisXu, Luqi; Zhang, Xinyuan
doi: 10.1117/12.3030084pmid: N/A
The aviation industry is a vital part of modern global travel; it is obvious that aviation review is worthy of further study. In our study, we mainly intend to explore whether we can use sentiment representation of reviews instead of the raw sentences, to assist the evaluation of customer satisfaction. Two sets of experiments were conducted, one with the reviews and one with their sentiment representation. Both sets employ deep learning and machine learning techniques to guarantee comprehensive study. Results show that traditional machine learning models achieved more competitive performances than deep learning models in two tasks, and the model with Gradient Boost using sentiment representation gets the best performance. The study finds that sentiment representation could serve as a viable raw material substitute further showing that a simplified approach can be used to achieve efficiencies without sacrificing accuracy for practical applications. This provides a solid reference for future studies that intend to develop fast and accurate classification models for airline reviews.
Front Matter: Volume 13162doi: 10.1117/12.3032702pmid: N/A
This PDF file contains the front matter associated with SPIE Proceedings Volume 13162, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.
Chess piece recognition using deep convolutional neural networksPapadimitriou, Orestis; Kanavos, Athanasios; Maragoudakis, Manolis; Gerogiannis, Vassilis C.
doi: 10.1117/12.3030405pmid: N/A
Chess piece recognition poses a significant challenge in computer vision due to the complex visual patterns and occlusions involved in identifying each piece’s type. In recent years, deep learning, particularly convolutional neural networks (CNNs), has emerged as a promising approach for image recognition, achieving state-of-the-art performance across various visual recognition tasks. In this paper, we propose a CNN-based approach for accurate chess piece recognition, capable of identifying the type of chess piece on each square of a chessboard. Our approach utilizes a deep neural network architecture that combines convolutional and fully connected layers to extract relevant features from chessboard images and make precise predictions. To evaluate our approach, we employ a large and diverse dataset of labeled chessboard images and compare its performance against state-of-the-art methods for chess piece recognition. Experimental results demonstrate that our approach surpasses existing methods, achieving an impressive accuracy of 98.9% on the test dataset.
Two-stage dual-channel driving distraction behavior recognition algorithm based on key point detectionCao, Wenjun; Chen, Shuaichao; Yu, Li
doi: 10.1117/12.3030001pmid: N/A
As an important part of automobile safety system, distracted driving behavior recognition has important research value. By analyzing the limitations and difficulties of the existing distraction driving recognition methods, this paper proposes a two-stage dual-channel recognition network. In the first stage, the Alphapose key point detection network based on SF3D data set pre-training is used to obtain the driver 's key point information, and the key area heat map is generated based on the Gaussian heat map. It is combined with the original image to form the two-channel input of the second stage. The fusion feature is generated by the feature fusion module based on feature concatenation, and it is used as the input of the second stage ResNet-50 backbone recognition network for recognition. Finally, in order to enhance the recognition effect, this paper introduces spatial and channel attention mechanisms to enhance the learning of interest features. And comparison and ablation experiments are designed for the proposed method. Compared with the benchmark network model, the proposed method improves 2.6 points, which verifies the effectiveness of the algorithm.
Advancing skin cancer classification across multiple scales with attention-weighted transformersYang, Guang; Luo, Suhuai; Li, Jiaming
doi: 10.1117/12.3030006pmid: N/A
Skin cancer remains a pressing global health concern, with millions of cases diagnosed annually. Early detection is vital, as survival rates vary significantly depending on the stage of diagnosis. Recent advances in dermatological practice have embraced dermoscopy, but its effectiveness often relies on practitioner experience, leading to diagnostic inconsistencies. In the realm of skin cancer classification, traditional machine learning methods gave way to deep learning, with vision transformers gaining prominence. This paper introduces a novel approach that leverages attention-weighted transformers for skin tumor classification. Attention weights gauge the significance of image patches, enabling precise region attention. Within our proposed framework, we introduce an enhanced transformer structure that capitalizes on the power of self-attention mechanisms. This architecture acquires discriminative region attention across multiple scales, enabling the model to effectively capture intricate image details and patterns. Experimental validation compares our method against Inception ResNet with soft attention and ViT-Base on the HAM10000 dataset. Data preparation involves duplicate removal, class rebalancing, and pixel-level augmentation. Evaluation metrics encompass accuracy, precision, sensitivity, specificity, and the F1 score. Results show our approach outperforms existing methods, achieving an accuracy of 93.75%. This work represents a significant stride toward accurate skin tumor classification, marrying innovative architecture with meticulous dataset preparation. The proposed approach holds potential to advance diagnostic tools for skin cancer, benefiting medical practitioners and patients alike.
Federated unlearning for medical image analysisZhong, Yuyao
doi: 10.1117/12.3030004pmid: N/A
Recently, federated learning has gained significant attention for its ability to train models without centralizing clients’ data on a central server. This unique characteristic makes federated learning widely applicable in medical image analysis, a field where ensuring patients’ privacy is imperative for medical institutions. However, in compliance with privacy regulations in certain regions, medical institutions must mitigate the influence of their clients’ data on the global model. Existing machine unlearning methods cannot be straightforwardly applied in this scenario, as they require access to clients’ data. Therefore, federated unlearning becomes a necessary solution. The basic strategies of federated unlearning are excessively time-consuming to be practical, prompting an urgent need for a more cost-effective approach. While previous works have proposed various strategies, they often prove either too costly or unstable for real-world applicability. In this paper, we adopt an approach called importance-based selection based on FedEraser, which expedites the retraining process at the expense of storage space. We also attempt to enhance its storage efficiency by pruning less significant updates. We conducted experiments on two datasets in medical image analysis, and the results vividly demonstrate the effectiveness of removing the target client’s impacts. The time and storage consumption of our strategy are also consistent with expectations, emphasizing its practicality.