Compression Techniques for Handwritten Digit Recognition
Hassan Alsobaie, Irfan Ahmad · 2020
Compressing images before recognition leads to many benefits including efficient computation, compact models, and optimal memory utilization. Several techniques for compression of handwritten digits have been investigated and implemented. This paper presents three compression techniques used in signal processing for compressing handwritten digit images, which are Discrete Cosine Transform (DCT), Discrete Sine Transform (DST) and Wavelet Transform (WT). These techniques are evaluated for their ability to compress the digit images while retaining useful information needed for classifying them, subsequently. Experiments conducted on the publicly available MINST dataset show the effectiveness of the techniques. With the presented techniques, we were able to compress the original images by 48.98%, 71.30%, and 87.24% while leading to reduction in accuracy by only 1.413%, 3.187%, and 7.238%, respectively, on an independent test set.