Bilateral Filtering NIN Network for Image Classification

Dong, Jiwen; Gao, Yunxing; Li, Hengjian; Guo, Tong

doi:10.1007/978-3-319-63312-1_58

Cited by 1 publication

(1 citation statement)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The convolutional layer outputs K (total character classes plus blank) channel feature maps, to generate one feature map for each corresponding category. A convolutional layer is preferred over the fully connected layer because it is more natural for the convolutional structure to enforce the correspondences between the feature maps and the categories [34]. The kernel size, stride, and padding size of the convolutional layer are set to 3 × 3, 1 × 1, and 1 × 1, respectively.…”

Section: B Temporal Mappermentioning

confidence: 99%

Fully Convolutional Sequence Recognition Network for Water Meter Number Reading

et al. 2019

View full text Add to dashboard Cite

One of the most widely used frameworks for image-based sequence recognition is the convolutional recurrent neural network, which uses a convolutional neural network (CNN) for feature extraction and a recurrent neural network (RNN) for sequence modeling. However, the RNN is computationally expensive in both training and inference, which limits its application in time-constrained systems. Some models replace the RNN with an attention mechanism for sequence modeling but still, require expensive iterative computations. In this paper, we argue that a CNN with sufficient depth can capture contextual information, eliminate the need for recurrent operations and thus be fully parallelized. We focus on the problem of water meter number reading (WNR), which is a typical sequence recognition task but has rarely been investigated. We propose a fully convolutional sequence recognition network (FCSRN) for the fast and accurate reading of water meter numbers. Furthermore, we design an augmented loss (AugLoss) function to manage the intermediate states of the digits and effectively improve performance. The experimental results demonstrate that the FCSRN has the ability to capture contextual information and eliminate the need for recurrent layers, and simultaneously requires fewer parameters and less computation. The FCSRN with AugLoss outperforms RNN-based and attention-based models. In addition, AugLoss can effectively improve the performance for RNN-based and attention-based models. Moreover, we constructed and released a dataset that contains 6000 water meter images with labels

show abstract

Section: B Temporal Mappermentioning

confidence: 99%