Variable illumination compensation for pigskin leather images using local and global block analysis

Variable illumination seems to be a main challenge to any automatic image processing system particularly in the segmentation and recognition part. The new intensity variations, which do not presence in the original image, may lead to a false segmentation result. In this research, we present a method to compensate the imbalance illumination in pigskin leather texture image. The main objective of this research is to offer an alternative pre-processing tool for further image enhancement procedures that is fast, simple and robust. The method utilizes the information within the local and global area of the image. The results are promising as the output image illumination condition is improved. In addition to that,the performances are either similar or better if to be compared to other established methods. However, the output images are affected by the visibility of the block boundaries. It happens due to the rapid intensity variation across the block boundaries, which is caused by the inaccurate residual selected for the normalization step. The next objective of this thesis is to remove the blockiness effect while still maintain the image attributes and improve the illumination condition. Some sets of modifications are employed to the basic model of local-global block analysis (B-LGBA) in order to eliminate the blockiness effect including horizontal improvement (HI-LGBA), vertical improvement (VI-LGBA), horizontal-vertical improvement (HVI-LGBA) and vertical-horizontal improvement (VHI-LGBA). The results produced by the enhanced LGBA illustrated that the modifications to be made are depending on the illumination trend. If the illumination is in one direction, the single modification is sufficient to remove all the boundaries. However, if the illumination is multi-directional, combined modifications are necessary, with the less-influencing direction must be dealt first. The accuracy of the proposed method is evaluated via the error of the numbers of the segmented region counted in the output image. Most of the outputs are oversegmented due to the existence of other low-intensity pixels which are contributed by the non-uniform surface of the sample and also the blocky pattern.From our 20 samples consist of different illumination types, the HVI -LGBA and VHI-LGBA yielded the lowest percentage of the errors with 29.52% and 35.43% compared to the errors produced by the B-LGBA which is equal to 120.35%.

Preview PDF FK 2012 23R.pdf Download (739kB) | Preview

Actions (login required)

View Item