Before handling classification, you should improve image quality. This matter includes: reduce noise, improve contrast etc. Another task might influence to classification is image compression. In this case, you should not compress image, or compress image with good enough compression ratio.

First of all, read the image and convert into grayscale and choose green channel as it performs better results, apply top hat transform for better background intensity, then apply global intensity threasholding or any other to eliminate noise (you can apply various approaches such as adaptive thresholding and Otsu's method for global intensity threasholding). Then perform your criteria for required output.

CNN generally does not require any preprocessing steps to be carried out on the input images. You can input the raw images with all RGB channels to the CNN. There are many references available which encourage to use images in RGB instead of any other color space in CNN. First try creating a suitable CNN model and the train it with raw images. If you are not satisfied with the accuracy then try with some conventional preprocessing steps as suggested by Azhar Imra. However, rarely the preprocessing steps do some significant improvement on the accuracy and in most of the time it worsen out the result.

Before handling classification, you should improve image quality. This matter includes: reduce noise, improve contrast etc. Another task might influence to classification is image compression. In this case, you should not compress image, or compress image with good enough compression ratio.