农业工程学报
農業工程學報
농업공정학보
2014年
14期
154-162
,共9页
李文勇%李明%陈梅香%钱建平%孙传恒%杜尚丰
李文勇%李明%陳梅香%錢建平%孫傳恆%杜尚豐
리문용%리명%진매향%전건평%손전항%두상봉
机器视觉%图像处理%特征提取%害虫分类%多类支持向量机
機器視覺%圖像處理%特徵提取%害蟲分類%多類支持嚮量機
궤기시각%도상처리%특정제취%해충분류%다류지지향량궤
computer vision%image processing%feature extraction%pest classification%multi-class support vector machine
由于野外诱捕害虫的姿态存在多样性和不确定性,使得利用机器视觉进行害虫的自动识别与计数仍然是一个难题。该文提出一种基于颜色和纹理等与形态无关的特征相结合和利用多类支持向量机分类器的多姿态害虫分类方法。通过对目标害虫图像进行不同颜色空间特征、基于统计方法的纹理特征和基于小波的纹理特征的提取,构建了6组不同组合的特征向量。将10阶交叉验证的识别率作为适应度函数值,利用遗传算法对各组特征向量进行降维筛选。最后利用基于有向无环图多类支持向量机分类器对多姿态害虫进行识别和特征组选择。结果表明,遗传算法最多可以使特征向量维数降到原来的38.89%,基于HSV三通道颜色图像的小波纹理特征组在建模时间和平均准确率方面都表现最优,可以作为一种有效的多姿态害虫分类特征选择。
由于野外誘捕害蟲的姿態存在多樣性和不確定性,使得利用機器視覺進行害蟲的自動識彆與計數仍然是一箇難題。該文提齣一種基于顏色和紋理等與形態無關的特徵相結閤和利用多類支持嚮量機分類器的多姿態害蟲分類方法。通過對目標害蟲圖像進行不同顏色空間特徵、基于統計方法的紋理特徵和基于小波的紋理特徵的提取,構建瞭6組不同組閤的特徵嚮量。將10階交扠驗證的識彆率作為適應度函數值,利用遺傳算法對各組特徵嚮量進行降維篩選。最後利用基于有嚮無環圖多類支持嚮量機分類器對多姿態害蟲進行識彆和特徵組選擇。結果錶明,遺傳算法最多可以使特徵嚮量維數降到原來的38.89%,基于HSV三通道顏色圖像的小波紋理特徵組在建模時間和平均準確率方麵都錶現最優,可以作為一種有效的多姿態害蟲分類特徵選擇。
유우야외유포해충적자태존재다양성화불학정성,사득이용궤기시각진행해충적자동식별여계수잉연시일개난제。해문제출일충기우안색화문리등여형태무관적특정상결합화이용다류지지향량궤분류기적다자태해충분류방법。통과대목표해충도상진행불동안색공간특정、기우통계방법적문리특정화기우소파적문리특정적제취,구건료6조불동조합적특정향량。장10계교차험증적식별솔작위괄응도함수치,이용유전산법대각조특정향량진행강유사선。최후이용기우유향무배도다류지지향량궤분류기대다자태해충진행식별화특정조선택。결과표명,유전산법최다가이사특정향량유수강도원래적38.89%,기우HSV삼통도안색도상적소파문리특정조재건모시간화평균준학솔방면도표현최우,가이작위일충유효적다자태해충분류특정선택。
Pest identification and classification is time-consuming work that requires expert knowledge for integrated pest management. Automation, including machine vision combined with pattern recognition, has achieved some applications in areas such as fruit sorting, robotic harvesting, and quality detection, etc. Automatic classification and counting of pests using machine vision is still a challenge because of variable and uncertain poses of trapped pests. Therefore, using Pseudaletia separata, Conogethes punctiferalis, Helicoverpa armigera, Agrotis ypsilon with different poses as research objects, this paper presents a novel classification method for multi-pose pests based on color and texture feature groups and using a multi-class support vector machine. 320 images were taken using field samples with an original resolution of 4 288×2 848. The subimages of pests with 640×640 pixel size were obtained from original images for computational efficiency. Color features in RGB and HSV spaces, statistical texture features, and wavelet-based texture features were extracted. Six feature vector groups were constructed using those features. In order to select effective feature parameters of each group, a genetic algorithm was designed to optimize feature vectors based on 10-fold cross-validation. Finally, the one-against-one DAGMSVM (acronym as yet undefined) algorithm was applied to classify and recognize the four kinds of target pests and to find the best feature group. 80 images (60 for the training set and 20 for the testing set) were adopted for each species. Parameter numbers were calculated and analyzed after optimization, thus the best parameters were selected for each group. The training time of the SVM model and classification accuracy, which contains false negative and false positive details, were compared between pre-optimization and post-optimization. The results showed that the highest parameter optimization ratio is from the sixth feature group with a dimension reduction rate of 61.11%. Compared with the RGB and statistical texture feature group F2, the optimization ratio of HSV and statistical texture feature group F3 is much better; that is, the latter one is more suitable to pest classification. Analysis and comparison between the optimization results of feature group F5 and F6 shows that the latter one is more suitable for multi-pose pest classification. The modeling time of each group has been greatly decreased, especially the one of group F6 (about 8 s), which is the minimum time of all groups with a decreased rate of 74.5%. Average accuracies of all groups have been improved beyond 97%. The sixth group has the highest accuracy (100%). Consequently, the sixth feature group, the feature vector of the wavelet filter in HSV color space, is an effective feature set for use in the classification of multi-pose pests. In addition, we have found that the feature parameters are similar among the misclassification pest sets, which may be improved by increasing the number of sample images in the training set.