Progressive structure network-based multiscale feature fusion for object detection in real-time application

Wang, Haifeng; Jiang, Lvjiyuan; Zhao, Qinglin; Li, Hao; Yan, Kai; Yang, Yang; Li, Songlin; Zhang, Yungang; Qiao, Lianliu; Cui-lian, FU; Yin, Hong; Hu, Yun; Yu, Haibin

doi:10.1016/j.engappai.2021.104486

Cited by 14 publications

(8 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Then, they fused the low-level deep and high-level semantic features into a similarity matrix. Wang et al [ 16 ] proposed a novel structure comprising three modules. One of the modules is responsible for multiscale feature alignment fusion.…”

Section: Related Workmentioning

confidence: 99%

Enhancing the Generalization for Text Classification through Fusion of Backward Features

Seng

2023

Sensors

View full text Add to dashboard Cite

Generalization has always been a keyword in deep learning. Pretrained models and domain adaptation technology have received widespread attention in solving the problem of generalization. They are all focused on finding features in data to improve the generalization ability and to prevent overfitting. Although they have achieved good results in various tasks, those models are unstable when classifying a sentence whose label is positive but still contains negative phrases. In this article, we analyzed the attention heat map of the benchmarks and found that previous models pay more attention to the phrase rather than to the semantic information of the whole sentence. Moreover, we proposed a method to scatter the attention away from opposite sentiment words to avoid a one-sided judgment. We designed a two-stream network and stacked the gradient reversal layer and feature projection layer within the auxiliary network. The gradient reversal layer can reverse the gradient of features in the training stage so that the parameters are optimized following the reversed gradient in the backpropagation stage. We utilized an auxiliary network to extract the backward features and then fed them into the main network to merge them with normal features extracted by the main network. We applied this method to the three baselines of TextCNN, BERT, and RoBERTa using sentiment analysis and sarcasm detection datasets. The results show that our method can improve the sentiment analysis datasets by 0.5% and the sarcasm detection datasets by 2.1%.

show abstract

Section: Related Workmentioning

confidence: 99%

Enhancing the Generalization for Text Classification through Fusion of Backward Features

Seng

2023

Sensors

View full text Add to dashboard Cite

show abstract

“…The recall calculates the proportion of adequately categorized positive patterns among all positive patterns. It is shown in Equation (8).…”

Section: Recallmentioning

confidence: 99%

“…In the majority of the scenarios mentioned above, rapid and accurate object detection in images is crucial. Deep learning algorithms are recognized for their excellent object identification and recognition ability 8 . When all things are considered for the practical applications that are indicated above, it is critical to arriving at efficiency in terms of the processing time of a single image and the accuracy in object recognition (objects might be either traffic signs or any other traffic‐related items) 9 .…”

Section: Introductionmentioning

confidence: 99%

“…Deep learning algorithms are recognized for their excellent object identification and recognition ability. 8 When all things are considered for the practical applications that are indicated above, it is critical to arriving at efficiency in terms of the processing time of a single image and the accuracy in object recognition (objects might be either traffic signs or any other traffic-related items). 9 This work considers and treats the challenge of detecting objects for recognizing Indian Cautionary Traffic Signs (ICTS).…”

mentioning

confidence: 99%

See 1 more Smart Citation

Recognizing the Indian Cautionary Traffic Signs usingGAN, Improved MaskR‐CNN, and Grab Cut

Satti

Devi

Srinivasan

2022

Concurrency and Computation

View full text Add to dashboard Cite

The detection and classification of traffic signs is a major challenge for self-driving vehicles. The task can be narrowed down to detecting and classifying Indian Cautionary Traffic Signs (ICTS). In this proposed work, the difficulty of detecting and identifying Indian Cautionary Traffic Signs is addressed, and sincere attempts have been made to attain a possible solution using a Generative Adversarial Network (GAN), Improved Mask R-CNN, and GrabCut algorithms. Data augmentation is done using Cascade Pyramid Generative Adversarial Network (CP-GAN) to upsize the data set. The Mask R-CNN with certain adoptions termed Improved mask RCNN is used in conjunction with the Grab-Cut method to handle ICTS detection and identification through automatic end-to-end learning. Initially, Improved Mask R-CNN generates a pixel-by-pixel segmentation mask for each item in the input sample. Masks developed using Improved Mask R-CNN are not always clean, that is, some background pixels are often seen in the foreground segmentation. Hence, the generated masks are refined using the Grab Cut algorithm to enhance image segmentation. This combined approach works well in isolating the traffic signs from the ground truth images. Improved Mask R-CNN attained better performance in the overall performance of traffic sign detection in the Indian data-set. This method recognizes 40 different types of cautionary traffic signs from the unique Indian data set. The results are provided for complicated traffic sign categories that have not been addressed before. The proposed technique is trained and evaluated on ICTS, GTSDB, STSD, LISA, and DITS data sets, and it is also compared against cutting-edge object recognition methods like Mask RCNN, Faster RCNN, SSD, and YOLOv3.

show abstract

“…At the current stage, there are four predominant methods for target detection based on SAR images: Target detection methods based on structural features [4][5][6] , grey-scale features [7][8][9] , texture features [10][11][12][13] , and deep learning [14][15][16][17] . In comparison, deep learning-based methods boast powerful feature extraction capabilities and are capable of automatically learning structured features to successfully achieve high-precision recognition of detection targets [18][19][20] . However, the research in the direction of SAR ship target detection still needs to overcome several severe tests, such as the influence of ship characteristics, radar characteristics, and environmental factors (cities, ports, islands, reefs, etc.)…”

Section: Introductionmentioning

confidence: 99%

Wavelet thresholding and F-NLM filtering based denoising algorithms applied to high resolution SAR ship detection

Tong

Liú

Peng

et al. 2023

Preprint

View full text Add to dashboard Cite

High resolution (HR) SAR ship images have distinctive features of multi-scale, multi-scene and densely arranged distribution of targets, while it is still challenging to have fast and accurate detection. To address the above problem, a SAR ship target detection framework using wavelet thresholding and Fast NLM (F-NLM) joint filtering is proposed. Firstly, wavelet thresholding and F-NLM are used to filter and de-noise the high-resolution SAR images to reduce background clutter noise, while enhancing the detection target detail features and edge information, solving the problem of high false alarm rate in multi-scale, inshore and offshore scenes of ships in SAR images of high resolutions.Then, the YOLOv5 detection network combined with the bi-directional feature fusion module (Bi-FPN) is selected to enable the model to better balance feature information. It can strengthen the aggregation of low and high semantic information and further improve the accuracy of the model. Experimental results show that the SAR ship detection framework has better robustness and target detection accuracy than other deep-learning based algorithms. Compared with SSDD and YOLOv5 network model, the Average Precision (AP) is improved by 1.33% and 2.25% respectively.

show abstract

Progressive structure network-based multiscale feature fusion for object detection in real-time application

Cited by 14 publications

References 25 publications

Enhancing the Generalization for Text Classification through Fusion of Backward Features

Enhancing the Generalization for Text Classification through Fusion of Backward Features

Recognizing the Indian Cautionary Traffic Signs usingGAN, Improved MaskR‐CNN, and Grab Cut

Wavelet thresholding and F-NLM filtering based denoising algorithms applied to high resolution SAR ship detection

Contact Info

Product

Resources

About