The Arsenal Algorithm: AI-Driven Weapon Recognition with CNN -SVM Model
Gaurav Gupta, Saumitra Chattopadhyay, Vinay Kukreja, Manisha Aeri, Shiva Mehta · 2024
This paper revolves around weapon recognition. It introduces the dual-framework dispute of the target to control two CNNs that are used to foreground jointly with the SVM-P and fuse the results. The system is rolled out and tested accordingly. The system is trained and evaluated using a dataset that includes five categories of weapons: firearms, edged weapons, explosives, improvised types of weapons, and chemical ones. We conduct stringent performance profiling of the Model based on various measurements, i.e., for each of the five types of weapons, namely, precision, recall, F1-score, and general accuracy. The Model's performance in recognizing armaments has been proven to be quite efficient, following the outcomes of our study. For the Class 1 weapons, the Model showed an accuracy of 89.86%, a recall of 97.01%, and an F1 score of 93.30%, with a remarkable total accuracy of 99%. Next, in class 2, the resilience score points towards a high resilience of the Model, which gives a 91.61% precision score, a 95.78%recall score, and a 93.65% f1 score. In addition, it recorded a final accuracy level of 98%. A macro average, calculated as a mean performance across all classes disregarding class imbalance, showed precision, recall, and mean values of 93.13%, 94.17%, and 93.60%, respectively. Considering the rate of occurrences inside each classification, the weighted mean brought somewhat particular viewing with precision, review, and F1-scores of 93.87%, 93.82%, and 93.81%, individually. In unbalanced datasets, the microaverage measures overall performance, and 93.82% was a constant value for accuracy, recall, and F1 score. The total accuracy of the Model was 93.82422803%.