Genetic Algorithm-driven Image Processing Pipeline for Classifying Three Bird Species: An Empirical Study of Two Encoding

This paper introduces iGRACE, a genetic algorithm designed to tackle the challenge of constructing and optimizing image processing pipelines, including hyperparameters. The pipelines encompass vital tasks such as noise filtering, image segmentation, feature extraction, and classification. The optimization process centers around minimizing the Cross Validation Error Rate (CVER) through the tailored selection of hyperparameters. In this study, the efficacy of iGRACE is examined using two distinct encodings: mixed and binary. While both encodings showcase comparable performance based on numerical metrics, the binary encoding outperforms the mixed encoding in terms of numerical results and execution time. However, what sets iGRACE apart is its unique attribute of providing interpretable and explainable solutions, particularly evident when comparing its results with those of a Convolutional Neural Network (CNN). Although no statistically significant differences emerge between the two encodings, a closer examination of the visual outcomes underscores iGRACE's strength in generating image processing pipelines that are more intuitive and comprehensible for endusers when employing the mixed encoding. Notably, this insight highlights the trade-off between numerical superiority and the interpretability advantage offered by iGRACE. Furthermore, there is room for improvement in the accuracy of numerical prediction values in the realm of classification. This signifies a potential avenue for future refinements to enhance the algorithm's predictive capabilities.