Research On Color Correction Method For Flue-Cured Tobacco Leaf Image Based On Ada-CatBoost Model

Color is one of the important characteristics of tobacco leaves,and the color of tobacco leaves can to some extent reflect their quality.Therefore,in the process of collecting tobacco leaf images,it is particularly important to ensure the authenticity of the color.However,in the actual sampling process,it is often affected by factors such as device sensors and environment,resulting in color distortion in the captured images.Therefore,conducting color correction research on distorted tobacco leaf images has important value and significance.At present,color correction models are often used to address the issue of image color distortion,but there are few studies that consider the differences in color distortion of pixels in different regions.The main work of this thesis is as follows:(1)Establish a dataset using evenly printed and flat color cards.In order to address the issue of uneven surface and color distribution of tobacco leaves,which made it difficult to accurately represent the true values of each leaf with a certain color value,a color card with relatively flat surface and uniform color was selected to establish an overall color correction dataset for model training.Given the difference in spectral reflectance between the color card and the surface of tobacco leaves,a mapping model for the two was obtained.(2)Establish an overall color correction model to address the issue of color distortion in tobacco leaf images.Transform the problem of image color correction into a numerical regression problem,selected commonly used regression methods for color correction,and then used the reciprocal variance method to combine the Ada Boost and Cat Boost regression tree models.The experimental results showed that the median color difference corrected by this method was 2.8,and the average color difference was3.24,which is the lowest among the methods used;The proportion of color blocks with higher structural similarity was 62.5%,and the proportion of color blocks with higher subjective evaluation scores was 82.3%,which was the highest among the methods used.Therefore,the model correction effect using the reciprocal variance method combination was the best.(3)Further achieve fine color correction for tobacco leaf images that have completed overall color correction.By analyzing the characteristics of the collection equipment and sensors,it can be seen that there are differences in the degree of color deviation between pixels.Therefore,a three channel difference model was established for each pixel to achieve fine color correction.The results indicated that about 90% of evaluators believe that the images that have completed fine color correction are consistent with the true color of tobacco leaves.In order to verify the feasibility and correction effect of the method used,multiple tobacco leaf images were selected for testing and validation.The results show that after completing the overall color correction of the image,further fine color correction of pixels can improve the color quality of tobacco leaf images.The accuracy of grading tobacco leaf images obtained from photography is 34%,and the accuracy of grading images that complete fine color correction is 78%,an increase of 44%.The method used has good practical value and has good application prospects.