Study Of Pear Surface Defects Detection Based On Laser Speckle Image

In the field of agriculture, laser speckle is a new non-destructive technology. When laser light with good coherence irradiate on the optically rough surface or the surface with some kind of activity, the scattering light will interfere with each other and form a mottled pattern called laser speckle. The laser speckle contains various information of the reflector, such as the roughness, the particle’s motion or the temperature information.We used Huangguan pear as our object in this research, focused on the identification of defects from stem/calyx and the sound area based on laser speckle image.(1) A laser speckle experiment platform with corresponding hardware and software was built up. The hardware system consisted of light box, laser light source, laser beam expander, polarizing film, color CCD camera, computer and digital signal generator. Software system combined with the hardware system which can realize the acquisition, storage, and processing of laser speckle image. Parameters could be adjusted, such as ROIs, frequency of external trigger source, the saving path and.etc.The image processing module mainly realized the speckle image preprocessing, including the RGB three-channel separation, ROI re-setting and.etc. Then the THSP and corresponding co-occurrence matrix could be constructed from the series of grayscale images, the inertia moment could be calculated based on the THSP. The method of Fujii, GD (Generalized differences) and WGD (weighted generalized differences) was implemented, which laid the foundation for subsequent testing.(2) Fujii method and WGD method were used to process the speckle of huangguan pear to get the result images of Fujii and WGD which showed different texture features representing different activity. A total of200pear samples were tested, and400sets of images were collected to extract texture features. The classification were carried out by single feature, and the results showed that the ASM (angular second moment) extracted from the result image of WGD had the best performance in distinguishing of the defects from stems and calyxes. The overall accuracy of calibration and validation were96.4%and96.7%respectively. Combinations of multiple texture features were tested in classification. The result showed that the combination of ASM and standard deviation of COR (correlation) had the best performance in classification, the overall accuracy of calibration and validation were both reached97.5%. The result proved that this method could be used in recognition of defects from stem/calyx.The influence of size of ROI on identification were studied, ROI with size of512×512,256×256,128×128,64×64,32×32and16×16were tested. With the comprehensive consideration of the classification accuracy and consumed resources, the512x512size of ROI were deemed to be the best size, and the combination of features (ASM_MEAN ASM_SDEV, COR_MEAN and COR_SDEV) extracted from WGD method had the best classification result, and the overall accuracy was98%.(3) The method of THSP was also used to analyze the laser speckle image. The THSP of defects, stem, calyx and sound area were built respectively. Then the co-occurrence matrix of THSP and the corresponding IM (inertia moment) were calculated. The different levels of activity showed different shape in the THSP and co-occurrence matrix. The co-occurrence matrix of defects showed more pixel distributed in the diagonal line, while co-occurrence matrix of sound part (stem/calyx/normal part) showed that more pixels spread in the whole image which looked like a piece of white cloud, which proved that the defect part had lower activity than the sound part.In order to illustrate this conclusion quantitatively, the IM values of co-occurrences of the THSPs were extracted, and the classification of samples were tested based on the IM values. The results showed that classifying of the sample only by the IM values was not good.The reason might be caused by manual building of the THSPs.