关键词: 鳞翅目, 图像识别, 颜色名, OpponentSIFT, 视觉词袋, 支持向量机

Abstract: 【Aim】 This study aims to realize the automatic insect image recognition by exploring the state-of-art computer vision technology. 【Methods】 The captured insect image was first preprocessed to remove the background and get the foreground mask. The minimum bounding box of the foreground was computed and the valid foreground region was cut out accordingly. The features on this valid region were extracted. The color name feature was extracted firstly. The pixels on original RGB (Red-Green-Blue) image were mapped to 11 color name planes, where the value represented the probability of the RGB value belonging to particular color name. Each color name plane was divided into blocks of 3×3 pixels. The average probability in each block was calculated and 11 values from 11 color name planes formed the descriptor of the center pixel in the grid. Finally, the bag-of-visualword features for color name descriptors were generated by histogram statistics on spatial pyramid. For OpponentSIFT (Opponent Scale Invariant FeatureTransform) feature extraction, the image was first transformed from RGB space to opponent color space, SIFT descriptors from different channels were extracted and concatenated, and then were pooled into OpponentSIFT bag-of-visual-words with histogram statistics on spatial pyramid. Two types of bag-of-visual-word features were concatenated into feature vector of the insect image. SVM (Support Vector Machine) classifiers were trained with the feature vectors extracted from training set and were used to recognize lepidopteran species through classification. 【Results】 The proposed method was tested in a database with 576 insect images and the recognition accuracy reached 100%. 【Conclusion】 The experimental results prove that the lepidopteran images can be recognized efficiently by using color name and OpponentSIFT features.

Key words: Lepidoptera, image recognition, color names, OpponentSIFT, bag-of-visual-word, Support Vector Machine (SVM)