Object recognition is an important computer vision and machine learning problem. A specific case is automatic target recognition (ATR) on radar images. ATR can be effectively used for border security, safety systems to identify either man-made objects (such as buildings, ground and air vehicles) or people, as well as for target surveillance. In other words, with ATR one can obtain any visual information about the ground and objects without direct physical contact.
In the project, our team has developed a custom classification algorithm based on two different tools.
We have been working with MSTAR dataset. This is a public dataset containing ten classes of vehicles with different orientations with 0.3mx0.3m.
Fig.1. MSTAR dataset with 10 vehicles
We have built two pipelines: SVM-based and CNN-based. For the former, we have constructed several feature sets, optimized parameters of the classifier using K-fold cross-validation and conducted comparative experiments.
Fig. 2. SVM application steps
The highest recognition accuracy of 94.1% was achieved using the fusion of vertical and horizontal image profiles. Such feature representation was around 20x times more compact than Haralick features, while provided better performance.
The second applied tool was CNN. Below is the proposed architecture:
Fig. 3. CNN architecture
Created CNN consists of 3 convolutional layers and 4 fully connected layers. Dropout layers were applied for careful control of overfitting.
Proposed CNN-based solution provided the state-of-the-art level of recognition accuracy.
Fig. 4. Classification results for CNN (left – confusion matrix, right – total accuracies)
The maximum achieved accuracy is 99.5%, which is the best result in the community for the public MSTAR database!