Weed image classification using Gabor wavelet and gradient field distribution

This paper presents an image analysis technique that utilizes a combination of a Gabor wavelet (GW) and gradient field distribution (GFD) techniques to extract a new set of feature vectors based on their directional texture properties for the classification of weed types. The feature extraction process involves the use of GW to enhance the directional feature of the images, followed by GFD implementation to produce histogram gradient orientation angles and additional steps to generate the histogram envelope. Next, curve fitting technique is used to estimate the envelope function to determine its quadratic polynomial equation, y = ax2 + bx + c and by taking the second derivative, the curvature value, a, is determined and used as a single input feature vector. The proposed technique was compared with another technique that also uses a single input feature obtained via GW algorithm implementation only. The overall classification accuracy utilizing the proposed technique is 94%, whereas using a GW only feature obtained 84% accuracy. The results obtained showed that this proposed technique is effective in performing weed classification.