Improved KNN based on wavelet analysis and identification of the red worm


Band-pass filter for filtering the signal decomposed into a series of bands on the analysis and processing. Images using wavelet high-frequency information extracted image texture features available, the image frequency domain wavelet decomposition can be four regions, namely LL, LH, HL and HH. LL sub-band for the low frequency component is similar to weight, concentrated most of the original image information. LH, HL, HH high frequency components, representing the original image details. Decomposition of each of LIJ wavelet decomposition can be again, and so on. 3-layer wavelet decomposition indicate Figure l (see page 146 Figure Fig.1). Let the image after wavelet decomposition,mulberry sale, its phase vertical, horizontal and diagonal directions of the wavelet coefficients are,, D, its energy is defined as: Ⅳ = [H (, y)] (2) Ivl Ⅳ = Σ [(, )] (3); lr = l hydrazine E: [D (,uggs sko,)]( 4) I, respectively, reflecting the energy of the first-order wavelet at the vertical, horizontal and diagonal directions of the image detail. erefore, the vector (,,)= 1,2 - effective blood worms and other plankton that the texture energy features. Bloodworm and Cyclops 3-layer wavelet decomposition image shown in Figure 2 (see diagram on page 147 Fig.2) is shown. Vector obtained by this method reflects the global features,nike free, these features do not take into account the spatial location of different details, so the performance of its lines of limited capacity. To solve this problem, first the image is divided into equal size and does not overlap the 32 × 32 pieces; then calculated for each piece of energy, these energies form a vector, the last of this vector normalized, the normalized After the vector is a wavelet-based energy features. 2 Ⅺ N classifier design 2.1 standard KNN classifier KNN algorithm is an instance-based classification algorithm. First, a test sample, calculate the concentration of training samples with the similarity of each sample, according to similarity to identify the most similar training samples. On this basis, then a sample for each class scoring, scores a training sample belonging to the class documentation and testing of samples and the similarity between documents. is sample was taken after completion of the class score statistics, that is sorted by score. To a reasonable classification, a threshold should be selected, you can break that test samples are more than the value of all classes. Specific algorithm is as follows: ① cording to a collection of items re-describe the characteristics of the training sample vectors; ② in the new samples arrive, according to the feature extraction methods to determine the new vector representation of the sample; ③ selected in the training sample with new sample most similar samples calculated as follows:. Σ. × 5 garlic kk where, d is the test sample feature vector,columbia outlet, d, for the first class of the center vector, M is the dimension of feature vectors for the vector of the first dimension. Generally used to determine the value of a first initial value, and then adjust according to the results of experimental determination of value, usually a few hundred to several thousand as the initial value; ④ test samples in a neighborhood, turn right to the same calculation for each class, that is, that the contribution of each dimension for classification is the same, this is not realistic, and the same weight makes eigenvector distance or cosine of the angle between the calculation is not accurate enough, thereby affecting the classification accuracy. In response to this lack of rights was proposed based on neural network re-calculations. In this method, the use of probabilistic methods for the preliminary CHI feature extraction and model aggregation, characterized by the weight calculation principle is: If a dimension in all of the low, if the value in each category are quite different, then it has strong classification ability, but the variance is just the state of the response variable distribution of the main indicators, this method is effective to improve the classification accuracy. In view of this, use Equation 6 to determine the quantitative characteristics of the sample of each correlation and classification by the classification function of a weight vector is given :·,, ...,) =, ≥ 0 (6) characteristics to determine the weight, you can modify the function of the distance between samples in order to better photos, access to different states, different attitude, different sizes of the various types of image samples, the choice in order to ensure the reliability of test results, the training sample set and the test samples do not support vector machine method training and classification test, the proportion of training samples and the average classification accuracy of the results shown in Table l ~ 3 (see table on page Tab1el l48-3). With the wavelet Fund National Natural Science Foundation (50778o48) (6o8o3O96); Natural Science Foundation of Heilongjiang Province (E200812); China Postdoctoral Science Foundation funded project (202). Author Zhaojing Ying (1978), female,bottes ugg, Jilin Yi Walter, MA, lecturer, engaged in pattern recognition, digital image processing research. Corresponding author. Received 20o96-22 Revised 20o98-l0