The production of sugar and ethanol in Brazil is very prominent economically and the reducing costs and improving the production system being necessary. The management crops operations of sugarcane and the control of weed is one of the processes that cause the greatest increase in production costs; because the competition that exists between cane plants and weed, for water, nutrients and sunlight is big, contribute to the loss of up to 20% of the useful cane. The use of image processing techniques has proven to be a tool to aid the decision, reducing production costs, because through the early recognition of infestation, it is possible to make the localized application of herbicides, reducing the impact on losses during cutting and harvesting of cane. Applying bag of words technique for recognizing weeds plants is proposed. The method is divided into three stages: vocabulary of visual words, training and classification. Were defined six varieties of weeds that have significant occurrence in cane fields infestation in the São Paulo State, which is the largest producer of sugar and ethanol in the country. The varieties of harmful plants chosen were: Panicum maximum, Euphorbia heterophylla; Brachiaria decumbens; Brachiaria plantaginea; Quamoclit Ipomoea; Ipomoea hederifolia. As main class was defined sugarcane (Saccharum officinalis). Digital images were obtained weekly between September and November 2013, using a digital camera (Nikkon Coolpix P510). The recognition of the images was developed in MATLAB R2012a. On classification stage was used the Support Vector Machine (SVM), which is a non-probabilistic binary classifier, being the methodology tested with a set of 105 images of seven kinds of plants (six weeds plants and sugarcane). The proposed method gotten average accuracy of 90.68% in the recognition, showing is more sensible in identification of plants Brachiaria plantaginea and Ipomoea hederifolia.