Automated inspection machines for wafer defects usually captured thousands of images on a large scale to preserve the detail of defect features. However, most transfer learning architecture requires smaller images as input images. Thus, proper compression is required to preserve the defect features whilst maintaining an acceptable classification accuracy. This paper reports on the effect of image compression using Fast Fourier Transformation and Discrete Wavelet Transformation on transfer learning wafer defect image classification. A total of 500 images with 5 classes with 4 defect classes and 1 non-defect class were split to 60:20:20 ratio for training, validating and testing using InceptionV3 and Logistic Regression classifier. However, the input images were compressed using Fast Fourier Transformation and Discrete Wavelet Transformation using 4 level decomposition and Debauchies 4 wavelet family. The images were compressed by 50%, 75%, 90%, 95%, and 99%. As a result, the Fast Fourier Transformation compression show an increase from 89% to 94% in classification accuracy up to 95% compression, while Discrete Wavelet Transformation shows consistent classification accuracy throughout albeit diminishing image quality. From the experiment, it can be concluded that FFT and DWT image compression can be a reliable method for image compression for grayscale image classification as the image memory space drop 56.1% while classification accuracy increased by 5.6% with 95% FFT compression and memory space drop 55.6% while classification accuracy increased 2.2% with 50% DWT compression.
Modifying fractal image compression technique by hybridation of crowding genetic algorithm and scattered method
