Abstract: Abstract: In order to realize non-destructive testing of moldy paddy during storage, the present study developed a computer vision system for laboratory analysis. Five kinds of fungi which mainly caused paddy mildew, including Aspergillus oryzae, Aspergillus nige, Aspergillus nidulans, Penicillium citrinum and Aspergillus versicolor, were used as research objects. Five fungi were cultured and prepared as suspension, which was then inoculated into paddy samples. Paddy was stored in the condition of 30 ℃ and 90% relative humidity to speed up the mildew. According to the mildewing degree, paddy was divided into 3 groups, i.e. control (no mildew), slight mildew and severe mildew. Computer vision system was used for image acquisition of 3 groups of paddy samples. A total of 120, 600 and 600 images of paddy samples were obtained for the groups of control, slight mildew and severe mildew, respectively. After image processing, gray scale, color in the color space of RGB (red, green, blue) and texture features (i.e., angular second moment, energy, contrast, entropy) were extracted using gray level co-occurrence matrix with a total of 68 parameters acquired. SVM (support vector machine) and PLS-DA (partial least squares - discriminant analysis) were used to build the discriminating models for paddy mildew and mildew type. To reduce the complexity of the model and the data redundancy, successive projections algorithm (SPA) was used to eliminate collinearity among the 68 characteristic variables. Then, 11, 13 and 14 optimal features were determined for the classification of moldy paddy, fungus type of slightly moldy paddy and fungus type of severely moldy paddy, respectively. The results showed that, using all the extracted features, SVM models could accurately distinguish between the control group and the mildew group of paddy, which got an overall classification accuracy of 99.7% and 98.4% for modeling and validation set, respectively; SVM models presented better distinguishing performance for paddy's severe mildew type than slight mildew type; concerning paddy's severe mildew type, the overall classification accuracy was 100% and 94% for modeling and validation set, respectively, and concerning paddy's slight mildew type, the overall classification accuracy reached 99.3% and 92% for modeling and validation set, respectively. As a whole, SVM model obtained higher accuracy than PLS-DA. Based on the preferred feature selected by SPA, SVM models still distinguished better than PLS-DA models for paddy's mildew. For modeling and validation set, the accuracies were respective 99.8% and 99.5% for the discrimination between no mildew and mildewing paddy, 99.8% and 90.5% for the discrimination among paddy's slight mildew type, and 100% and 95.0% for the discrimination among paddy's severe mildew type. Therefore, the computer vision technique is feasible for paddy's mildew detection; the preferred features determined by SPA can well reflect paddy mildewing features. Using the preferred features, SVM models are able to identify and distinguish paddy mildew with satisfactory results, which can provide technical support for practical application.