Weight estimation based on image processing using artificial neural network

Accurately and fast detection of weight of objects has an important place for lots of academic and industrial application at the present time. In this work, it was aimed to weight estimation in a partially distance and camera angle independent manner based on image processing using artificial neural networks (ANN). Estimation was realized with egg and orange samples which is different categories of food sector. The constituted system consists of a camera, artificial lighting system, reflector and reference image. Object weight was estimated by processing the acquired image. Feed forward multilayer perceptron (MLP) and radial basis function (RBF) was compared as artificial neural network with experiment. 250 samples which consist of 4 different brands and 4 different classes (x-large, large, medium, and small) were used for testing and training the system on egg study. 150 orange samples which were different sizes were used as to orange study. It was tried to be obtained optimum system by using different feature vectors, different images and different artificial neural network parameters. Each tries which was realized with changes was run 5 times and then system performance was evaluated by getting arithmetic mean. Performance of optimum results between 31 methods was calculated for MLP as correlation coefficient (R) was 0.9942 and mean absolute error (mae) was 0.80gr and calculated for RBF as correlation coefficient (R) was 0.9943 and mean absolute error (mae) was 0.81gr for egg samples. As accurate classification rate had been %47 according to defined by Turkish food codex egg communique and classification values written on the egg cartoon, this rate was increased to %90.5 for MLP and increased to %86.3 for RBF as a result of this work. Performance of optimum results between 32 methods was calculated for MLP as correlation coefficient (R) was 0.9820 and mean absolute error (mae) was 2.58gr and calculated for RBF as correlation coefficient (R) was 0.9825 and mean absolute error (mae) was 2.42gr for orange samples. All processes were realized by using Matlab program and toolboxes. In addition, all images which were obtained during these works have been shared with the intent of assisting next works. A database which contains right angled, positive angled and negative angled 750 egg images and right angled, positive angled and negative 450 orange images was created.