Project

The Bode Description Project implemented a random forest Machine Learning model from decision trees to identify and classify targets within an image. Correct development of this AI model will utilize an FFT to obtain a rotationally invariant feature within the model and allow for efficient usage of time and hardware while giving an accurate depiction within the model. 

A common problem among image analysis is how rotational variance affects the model. A car, airplane, etc. may be oriented any which way. Running every possible orientation costs RAM, computation, and time. The problem is to develop a model that has rotationally invariant qualities, such that the model may identify the same target at any other orientation. 

System

Methods

Polar to Cartesian Coordinate Conversion

A tree contains several layers to output a value dependent on the input values. The model employs several trees to form a forest. The output from these trees is put into a vector of scalar values to weigh the outputs and determine a positive or a negative detection. 

An FFT is then applied to give the model its rotationally invariant feature. 

Conclusion

The model takes 2.15 hours to scan an image—employing a parallel architecture over 36 CPUs. 

The chosen parameters were then determined to be 1.4 radii increments, 1.65 arc length increments, and a sub-image size of 25×25. This performs well while maintaining a precision and recall optimization.