Final Project for NIC
See the attached report pdf for further details and results.
Genetic Algorithm (GA) / Neural Network hybrid for image compression and optical digit recognition. The program uses a genetic algorithm to assign integer symbols to each bit of an image. A perceptron is then trained with the sum of bits of each symbol as the input using a training image set. Finally, the perceptron is tested on a test image set.
First make the project
Then run ./RK with the following parameters, in this order:
Population: [Positive integer]
Selection Type: [String]
"bs" - Botlzmann Selection
"ts" - Tournament Selection
"rs" - Rank Selection // This is the best one
Crossover Type: [String]
"uc" - Uniform Crossover
"1c" - 1-Point Crossover
"nc30" - 30-Point Crossover // Replace '30' with any positive integer for any N-Point crossover
Crossover Probability: [Double, between 0 and 1]
Mutation Probability: [Double, between 0 and 1]
Number of Generations/Iterations: [Positive integer]
Number of Symbols: [Positive integer]
Example command, pretty good results:
./RK 100 rs uc .5 .01 200 32
Our tests yielded two notable results
- Individuals with only 16 symbols were able to achieve over 80% digit recognition accuracy. This is significant because 8x8 grid compression, which effectively has 64 symbols, was unable to achieve over 70% accuracy. This method of compression gave more accurate results with 1/4 the number of perceptron inputs.
- The algorithm was able to achieve 93.2% accuracy with 32 symbols. In comparison, the peak accuracy for the perceptron with uncompressed 32x32 bitmap inputs was 94.6%. This algorithm was able to achieve comparable accuracy with 1000 fewer perceptron nodes.
- It is likely that even better results are possible if the parameters for the GA and perceptron are further optimized. Given more time and computing power, the parameter space should be explored more thoroughly. The mutation rate will be an important target for further optimization, as it enables the algorithm to escape local minima.
- Adding a hidden layer to the neural network may improve the fitness of our algorithm and make it possible to compress images with even fewer symbols while maintaining or improving the accuracy.