[논문리뷰] A Convolutional Neural Network Approachto the Classification of Engineering Models

Preprocessing for classification

A weighted Light Field Descriptor (LFD) scheme is chosen as the method of feature extraction, and the generated images are fed as inputs to the CNN [1]

각각의 3D 모델링 데이터의 정십이면체의 정점에 20 대의 카메라를 배치하여 다양한 뷰에서 3D 모델의 이미지들을 캡처한다.(20대는 [1] 기준)

이렇게 캡처된 이미지들은 일치하는 CAD model label에 맞게 assigned된다.

 

딥러닝 학습 시 3D CAD 데이터 augmenting

1. By observing the 3D objects in each category, an overview of the 3D designs is obtained.
2. Using this knowledge, a few representative models are parametrically designed in Autodesk Fusion360 software - for each class
3. Following this, more 3D models are generated via a python script linked to the Autodesk Fusion360 API, for various sets of parameter values

• For example, in order to model a cuboid parametrically, one needs three parameters - for the three dimensions (say l, b, h). For different sets of values for l, b & h, we get corresponding cuboids. This process essentially generates multiple variants of a certain category by using many sets of parameter values.
• 저자가 위의 방법을 사용해서 만든 dataset를 이용할 수 있는 깃헙링크 : GitHub - bharadwaj-manda/CADNET_Dataset: This is the repository for the 'CADNET' Dataset, associated with the paper "A Convolutional Neural Network Approach to the Classification of Engineering Models".

ResNet for classification

-ResNet 설명

• It uses the idea of residual connections, that helps in faster training of deeper networks.
• input of each layer is simply the output of the previous layer. the input of each layer will be a ->summation of the output of the previous layer and the value from the residual connection.

 

-ResNet에서 사용되는 이미지와 3D CAD model 데이터의 차이점

1. ResNet에 주로 사용되는 image데이터는 실제세계의 정보를 담고 있지만, CAD 데이터는CAD모델에서 추출한 이미지이므로, 실제 세계의 디테일한 이미지와 비교했을 때, 좀더 구체적이지 않다.
2. CAD 데이터는 RGB정보가 아니다.
3. 3D모델에서  256X256의 이미지를 추출한것이기 때문에, 이것은 outer boundary for the 3D objects.에 불과하다. 실제 모습을 온전하게 표현할 수 없다.

 

-ResNet 변형

1. The proposed network requires a fewer number of filters to capture the features in the initial layer (32) as opposed to ResNet (64).
2. We require a lesser number of filters (in the hidden layers) compared to ResNet to extract the features from the images. Hence we have six layers of 32, 64, 128, 256 and 512 (= 5952 filters) as opposed to 6 layers of 64, 8 layers of 128, 10 layers of 256 and 6 layers of 512 (= 7040 filters) in ResNet.
3. This, in turn, reduces the number of parameters (by about a million in the network)
4. The output from the last hidden layer is then fed into a fully connected layer, with the number of nodes equal to the number of classes. In our case, it is 43 for CADNET. The activation function used here is softmax

 

 

딥러닝 모델의 classification동작 그림

 

- using a class weights approach

• class imbalance  : 분류한 데이터셋의 class가 불균등할 때 나타나는 문제
• 해결방법 :  The problem of class imbalance in the dataset is addressed using a class weights approach
• We estimate class weights for the unbalanced dataset. The weight of each class is simply the ratio of the number of samples to the product of classes and the bin-count of the class labels.

 

-post preprocessing

• ResNet의 Output를 산출할 때, label된 2D images에서 올바른 output를 산출할 수 있어야 한다.
• the machine learning algorithms such as XGBoost and CatBoost are used to do this.
• some images are misclassified because models from different classes appear similar from a certain viewing direction.
• In order to reduce such misclassifications and thus enhance the prediction accuracy, we process the probability values using XGBoost and CatBoost.