Improved adaptive watershed method for segmentation of cotton leaf adhesion lesions

Accurate segmentation of cotton leaf adhesion lesions not only can extract the feature vector of individual lesions to a large extent, but also is important for the improvement of the accuracy of lesion identification. Some methods for adhesion lesions segmentation may result in over-segmentation or under-segmentation. In addition, some segmentation methods for adhesion lesions can better solve the segmentation problem, but the selection of parameters is sensitive, which needs to be manually set and cannot adapt to complex conditions, such as H-minima transform. Aiming at the adhesion problem between lesions, an adaptive watershed segmentation method was proposed. Firstly, for the cotton lesion area extraction, the Gaussian filter was used for image filtering processing, and then the super green color component was extracted and OTSU binary segmentation was performed. Post-segmentation processing was carried by the mathematical morphology of hole filling and the morphological opening operation, so that the contour of segmented lesion area was continuous and the edge was smooth. Secondly, the local minimum threshold (h) was determined. Based on the H-minima watershed segmentation method, the proposed method combined the least squares method error theory to fit the least squares of each connected component in the image. And the least squares error value was calculated, then the contour irregularity of each connected component was determined according to the least squares error value. The minimum threshold h of the H-minima transform was determined based on different contour irregularities. Finally, Watershed segmentation of cotton leaf adhesion lesions was achieved based on different minimum thresholds. A total of 160 images with 2 to 5 adhesion lesions were selected from the lesion sample library for testing. The lesions segmentation accuracy for cotton leaves with different number of adhesion lesions was 91.25% with running time of 0.088 s. The proposal method achieved the automatic segmentation , and was especially suitable for the adhesion of different lesions and irregularities. Meanwhile, 150 samples with high degree adhesion lesion of ring disease, brown spot, anthracnose, leaf spot and cotton boll blight were selected for different segmentation methods contrast test. The results of contrast test - showed that the proposed method could automatically segment the adhesion lesions of 5 diseases. Distance watershed method, Gradient watershed segmentation method, marker watershed segmentation method, Chan-Vese method, Gaussian mixture method and the proposed method, the correct segmentation ratio were 67.8%, 36.4%, 83.7%, 70.3%, 82.1% and 93.5%, respectively, with the average running time of 0.034, 0.036, 0.046, 0.357, 0.108 and 0.094 s. The experimental results showed that the proposed method was superior to the other 5 methods, especially the over-segmentation problem was effectively suppressed. The proposed method took slightly longer time than the distance watershed segmentation method, the gradient watershed segmentation method, and the marker watershed segmentation method, and was lower than the Chan-Vese method and the Gaussian mixture method, which could still meet the real-time image processing requirements. The results of lesion segmentation test in complex environment showed that under complex conditions such as complex background, uneven illumination and uniform lesion size, the proposed method could better achieve the segmentation of adhesion lesions, and its segmentation accuracy and running time could meet the actual needs. The proposed method can not only automatically segment the adhesion lesions on cotton leaf, but also provide reference for the segmentation of adhesion lesions for other crop leaves.