In view of the difficulty in region location and classification of potato typical diseases under natural conditions, a new adaptive features fusion and fast recognition method of potato typical disease images was proposed. The segmented disease image, processing object of the proposed method, could be obtained as following two steps. Firstly, by using K-means, Hough transform and superpixels segmentation algorithms, the whole potato blade containing disease region was located in complicated background. Secondly, the disease region was separated from green blade by combining with two-dimensional Otsu and morphology method. On the basis of the segmented disease image, totally 124 potato disease features, including 18 color features, 21 shape features and 85 texture features were extracted. As thus, the color, shape and texture features were fused adaptively based on principal component analysis (PCA) algorithm and weighted formulation, and used to potato diseases recognition by support vector machine (SVM). According to features fusion and SVM recognition, totally 13 weighted principal components were gained as following three steps. Firstly, color, shape and texture features were automatically divided into many feature blocks, including RGB and HSV, geometric statistics (GS), central moments and Hu moments, Gray-level co-occurrence matrix (GLCM), high frequency low order moments and low frequency low order moments (HMLM), and high frequency covariance matrix eigenvalues and low frequency lower order moments (HELM). By comparison of recognition rates and features dimension, RGB, GS and HELM feature blocks were selected from color, shape, texture feature blocks, respectively. Secondly, five RGB, five GS and three HELM principal components were acquired by PCA algorithm. Thirdly, RGB, GS and HELM were weighted based on their recognition rates of principal components, and each principal component was also weighted based on weight distribution formulation. The recognition test of three kinds of typical potato samples showed that the proposed method had an obvious advantage. By using the same SVM recognition model, and compared with recognition rates of traditional adaptive methods, including PCA descending dimension, feature sorting selection, and so on, the proposed adaptive feature fusion algorithm had high average recognition rate which was increased by at least 1.8 percentage points. By using the same 13 adaptive fusion features, average recognition rate of the proposed recognition method was 95.2%, which were increased by 3.8 percentage points and 8.5 percentage points than those of ANN and Bayes, respectively, and run time of the proposed recognition method was 0.600s, which was 3s faster than that of ANN. Therefore, the proposed method could be used to greatly improve the recognition speed based on effectively ensuring the recognition accuracy.