Adaptive Probabilistic PCA Method on Color Image Inpainting and Its Application in Plant Leaf

Because of the influence of nature meteorological condition and background environment during the acquisition of the plant leaf image, the image degradation is always unavoidable with the salt and pepper noise. The image of plant leaf is generally characterized by rich textures and well-defined edges. It is unfavorable to the subsequent processing of color image with noise pollution. Although there are several filtering methods such as average filtering, wiener filtering, gauss filtering and median filtering, they do not satisfy the requirment on effective repairation and texture reservation of image. Consequently, to repair the image successfully with the textural details preserved and the edges clear, a new model for color image inpainting was proposed and called adaptive probabilistic PCA method. The procedure of the proposed model included 2 steps.After the leaf vein was identified and tracked based on tree, the vein inpainting was conducted by the probabilistic principal component analysis (PPCA) model, in which the iterations were adaptively selected according to the PSNR value of the restored images. To evaluate the effectiveness of the proposed model, a 3-step simulation test was invloved, and the evaluation criteria based on SNR and structural similarity image measurement(SSIM) was used to measure the degree of image distortion and similarity between the processed and the original image. Firstly, to determine the optimal iterations of the PPCA model, the inpainting results in different iterations were compared. Secondly, to test the image inpainting ability, the polluted images are simulated with different levels of noise. Finally, the proposed model had some comparison with the conventional filtering methods. The experiments showed that the iterations about 550 were appropriate while using the PPCA model for image inpainting. The restored image obtained by the proposed model was less residual noise and clearer textures than other filtering methods visually. The PSNR value of restored image was 26.819 9 dB, which was higher than using the wiener filtering, gauss filtering, average filtering and median filtering, by 9 dB, 7 dB, 6 dB and 1 dB, respectively. It was higher than the PSNR value of the noisy image by 14.48 dB. The SSIM value of restored image was 0.955 7, which was the largest among the above-mentioned methods. It indicated that the restored image using the proposed model was closer to the original image in the brightness, contrast and structure aspects. It could provide technical support to the subsequent processing of the color image. © 2017, Chinese Society of Agricultural Machinery. All right reserved.