Seyed Mehdi Yavari; Zahra Azizi
Abstract
Extended AbstractIntroductionLack of uniform light radiation on the objects, reduces the amount of contrast in the images and makes it difficult to extract image features. This problem destroys information about the behavior, shape, size, pattern, texture, and tone of the effects, and compresses the ...
Read More
Extended AbstractIntroductionLack of uniform light radiation on the objects, reduces the amount of contrast in the images and makes it difficult to extract image features. This problem destroys information about the behavior, shape, size, pattern, texture, and tone of the effects, and compresses the image histogram in one or more specific areas. UAV images have been widely used in recent years due to their extensive coverage, high operating speed, use in hard-to-reach areas and up-to-date equipment. If drone images are correctly taken and pre-processed, they provide good accuracy for a variety of applications. The preprocessing is important since the image acquisition conditions cannot be changed in most cases so that the acquired images are contaminated with some distortions or errors which must be removed or their effect reduced to a minimum before any process. Improving the exposure in the image, which increases the amplitude of the histogram, can highlight features with similar gray-scale values, and this is useful in identification. Materials & MethodsIn this study, two aerial images have been used with a variety of vegetation, soil and man-made features using Storm 2 hexacopter drone in Simorgh city (Kiakla) in Mazandaran province with longitude and latitude 52⸰ 54' 1'' and 36⸰ 35' 49''. At first the SMQT algorithm is applied to the input images. So the bits number of the input image is calculated to determine the number of transmission levels. Then with rgb2gray command creates a gray image of the original image. The overall average of the image is calculated and the DN of each pixel is compared to the average. If the DN is greater than the pixel value, the number 1 is assigned to the pixel, otherwise the number zero in another image is assigned to the pixel. The average calculation and segmentation of pixels based on the number of bits continues, each segmentation is called a transfer. Then, by converting the data from these divisions into values in the spectral range of the image, a new image is created. This image has higher radiometric resolution than the original input image but lower spectral resolution. For this reason, the image is fused. Global gamma correction is applied to the fused image. Finding gamma in the image, especially local gamma is time consuming and complex for programming and computing. Therefore, to increase the computing speed, a local gamma of 0.7 was applied to the whole image and then the first step processes are applied again and finally, the SSIM index is checked for image enhancement.Results & DiscussionThe SSIM value for input image 1 and 2 is 0.8372 and 0.8401 while this value before processing was 0.4352 and 0.4161. Examining the histogram of the images before and after processing, in all three bands R, G and B, shows the stretch of the image histogram in the range of 0 to 255. There is a decrease in the number of peaks and valleys in the histogram of the processed images. The density function for input and processed images shows that the more homogeneous the number of effects in the image, the greater the slope of the function graph. The value of the density function has increased after processing, which is due to the stretching of the image histogram. SSIM is used to validate the results in this study. The images have been visually improved significantly, but this is not enough for verification. The goal of quantitative quality recognition is to design computational methods that can accurately and automatically express image quality, which affects all the image pixels in the same way. The SSIM range is between (+1 and 0). The closer the measured value for an image to one, the better image quality will be. SMQT also has less computational complexity and less configuration. If the image of a light object is formed in a completely dark background (such as night shooting), this algorithm does not work in the background pixels. Examining the image samples taken from a complication at night, it was found that the black pixels changed color to purple after fusion. In order to optimize the algorithm, it is suggested to increase the efficiency of the algorithm by examining the spectral behavior of different features in different color spaces and integrating their effective components in image or feature highlighting or the use of plant or soil indicators. The fuzzy method can also be used for semi-shady areas. These improvements should also prevent complexity of computing by increasing efficiency.
