Can a Convolutional Neural Network Implement Histogram Equalization in Image Analysis? Open Access
Downloadable ContentDownload PDF
Recently, the Convolutional Neural Network (CNN) as a typical deep learning technique has been widely used in many image processing applications such as classification, recognition, segmentation and image reconstruction and achieved remarkable success. We have applied CNN to segment the breast region from thermal breast images. In that study, an interesting phenomenon showed that: image pre-processing by contrast-limited adaptive histogram equalization (CLAHE) can improve segmentation outcomes. The benefit of pre-processing is natural in traditional image processing, but it is questionable in deep learning. In a CNN, if any pre-processing could improve the final result, it will be automatically learned and implemented by the initial convolutional layers. That is one key feature of CNN. Thus, a likely reason for this phenomenon is: CNN cannot learn and implement histogram equalization (HE) to images. That is likely the case because convolutional operations in CNN are localized, but HE is a global process, and thus a CNN may inherently be unable to perform HE on images. To test this hypothesis, we firstly designed a simple 3-layer CNN and its input and output are same-size images. Second, the training images for CNN are from A category (thermal breast images) and the supervisors are histogram-equalized images. Third, we tested the trained CNN by different images from A category and some images from B category (portrait and nature images), which has distinct difference from A category. Then, we quantitatively compared the output images from CNN with histogram-equalized testing images via the Cumulative Histogram Similarity (CHS) between two images. Results show that CNN trained on A category performed about 40% lower on average CHS than HE to process B category. Such preliminary conclusion demonstrates CNN does not well learn HE but somewhat image style transformation to a certain category. Because once HE is learned, it should be implemented as needed for any kind of images. Our next studies are to examine the hypothesis on more categories' images and varied CNN architectures. In addition, we will seek a theoretical explanation of such limitations of the CNN. Two major points of significance of this study are: 1) it is a good way to define some of the limitations of CNNs, and thus helps understand them further; 2) if our hypothesis is eventually confirmed, we will examine CNNs in combination with other image pre-processing approaches. Such conclusions could encourage more CNN applications to use image pre-processing to improve their performance.