Page 72 - Read Online
P. 72

Ji et al. Intell Robot 2021;1(2):151-75  https://dx.doi.org/10.20517/ir.2021.14     Page 165

               Table 3. Adoption of deep learning models
                Deep learning model                                       Ref.
                AlexNet, ResNet                                           [76]
                Autoencoder                                               [89,101]
                CNN                                                       [61,62,65,68,69,74,75,79,85,94,95,102,108,109]
                CNN (object detection), RNN (distance estimation)         [93]
                CNN + YOLOv3                                              [117]
                CNN based self-proposed DFF-Net                           [118]
                CNN based self-proposed FR-Net                            [72]
                CNN to extract feature, only 1 class                      [82]
                CNN, transfer learning                                    [64]
                CNN, transfer learning, Bayesian optimization to tune hyperparameters  [100]
                CNN-LSTM                                                  [87,103,107]
                Faster R-CNN                                              [78,88]
                Faster R-CNN + CNN                                        [73]
                FastNet, convolutional network-based                      [120]
                Fine-grained bilinear CNNs model                          [70]
                FCN                                                       [119]
                GAN for CNN                                               [115]
                Inception-ResNet-v2 & CNN                                 [113]
                LSTM-RNN                                                  [63,71,99]
                Mask R-CNN                                                [121]
                ML Tree based methods                                     [80]
                MobileNetV2, YOLOv3                                       [84]
                Multilayer feedforward neural networks based on multi-valued neurons (MLMVN)  [60]
                neural network                                            [96]
                Point Cloud deep learning                                 [92]
                ResNet classifier, DenseNet classifier                    [81]
                ResNet, FCN                                               [83]
                Resnet50, transfer learning, Inception, Faster R-CNN      [67]
                Self-proposed, 2 stage FaultyNet, CNN based               [97]
                Self-proposed, segment U-Net (CNN based) then detect, progressive  [116]
                Self-proposed, ShuffleNet-v2 extracts features from the track image, RPN predicts  [86]
                Siamese convolutional neural network                      [66,91]
                Single Shot multibox Detector (SSD)                       [90]
                SqueezeNet, MobileNetV2                                   [111]
                U-Net to segment                                          [105]
                Variational autoencoder                                   [98]
                YOLO V3                                                   [77,104,106,110,114]
                YOLOv5 detect object; mast R-CNN detect surface defect; ResNet classify fastener state  [112]

               CNN: Convolutional neural network; RNN: recurrent neural network; YOLOv3: You Only Look Once, Version 3; LSTM: long-short-term memory;
               DFF-net: differential feature fusion convolution neural network; FR-net: feature fusion refine neural network; FCN: fully convolutional networks;
               RPN: region proposal network.


               randomly assigned for training and testing purposes. Fourth, the selected deep learning model is trained
               with the training data and validated by the testing data. Depending on the purposes, the deep learning
               model could perform classification or localization tasks. It is also possible to perform classification and
               localization concurrently, which is the most common type of task. The feature representations of the input
               data are always extracted; however, the next steps to deal with the feature vectors differ. It is noted that
               researchers also proposed their own neural network architectures to replace or complement the existing
   67   68   69   70   71   72   73   74   75   76   77