Military Equipment Classification

Model by Modzy

This model accepts JPEG images of any size and can classify the images into one of 87 military equipment classes. The model returns its top 5 predictions with their confidence scores, along with the explainability object for the respective predictions, to the user in a JSON file. This model can be used in multiple ways, such as classifying a large set of images into different military equipment categories based on object classes presented in the images.

  • Description

    Product Description

    PERFORMANCE METRICS:

    Explainable: This model has a built-in explainability feature. Click here to read more about model explainability.

    Adversarial Defense: This model has a built-in adversarial defense feature. Click here to read more about Modzy’s proprietary adversarial defense strategy.

    64% Top 1 Accuracy

    85% Top 5 Accuracy

    This model was trained on the Jane’s dataset. The Jane’s training dataset is a large imagery dataset of 64,906 images and 87 different object classes. The model was tested on the Jane’s test dataset of 16,227 images using 5-fold cross-validation. The model obtains top 1 accuracy of 0.64 and top 5 accuracy of 0.85 on the test dataset. Additionally, this model obtains a state-of-the-art top 1 robustness accuracy of 0.55 and top 5 robustness accuracy of 0.81 on the synthetically generated adversarial test dataset of 16,227 adversarial images generated by a range of different adversarial attacks and attack strengths. Some of this model’s strengths include its robustness against white-box and black-box adversarial attacks, and its capability in detecting and classifying object classes for a large range of images of different sizes and pixel distributions with the capability of finding the top 5 appropriate object classes that the image may belong to. For fast inference time, this model works best on a GPU.

    Top 1 Accuracy: The ratio of the number of correct predictions of the top 1 predicted class to the total number of input samples. Further information here.

    Top 5 Accuracy: The fraction of the top 5 predictions made by the classifier. Further information here.

    OVERVIEW:

    This model utilizes a residual neural network called ResNet50, which was first developed and published by Microsoft Research. ResNet50 uses residual learning which leads to networks that are very deep and learn a large set of features from the training dataset. The features learned by this deep learning model are low-level, mid-level and high-level features presented at the end of each sub-block inside the network. These features are residual features that can be simply understood as subtraction of features learned by the previous layer. We have trained this model using our Lyapunov based robust training solution. The features learning by this model are not only representative of the object classes in images but are also robust against adversarial attacks. This model was developed using the TensorFlow deep learning framework.

    TRAINING:

    This model was trained on the Jane’s dataset. The Jane’s training dataset is a large imagery dataset of 64,906 images and 87 different object classes. The training took 3 days on one DGX-1 GPU. The model was trained using stochastic gradient descent with a learning rate of 0.01 and momentum of 0.9 for 100 epochs on 64,906 training images.

    VALIDATION:

    The performance of the model was tested on a validation dataset of 16,227 images using 5-fold cross-validation. The model obtains top 1 accuracy of 0.64 and top 5 accuracy of 0.85 on the test dataset. Additionally, this model obtains the state-of-the-art top 1 robustness accuracy of 0.55 and top 5 robustness accuracy of 0.81 on the synthetically generated adversarial test dataset of 16,227 adversarial images generated by a range of different adversarial attacks and attack strengths.

    INPUT SPECIFICATION

    The input(s) to this model must adhere to the following specifications:

    Filename Maximum Size Accepted Format(s)
    image 50M .jpg, .png

    OUTPUT DETAILS

    This model will output the following:

    Filename Maximum Size Format
    results.json 1M .json