Identification of noisy poultry portion images using a neural network

The automatic sorting of products in food processing plants has become more important as faster processing systems emerge continuously. In raw poultry processing plants, popular chicken portions such as breasts, drumsticks, fillet, legs, and wings have to be sorted prior to packaging. Some current sorting methods use the portion weight as an indicator, however varying sizes of the different portions could result in similar weight amongst the different portions; thus causing incorrect identification. One solution is using portion images and an intelligent classifier to identify the portions, therefore, simulating the way human laborers sort the portion in a production line. The problem with the intelligent imaging method is the potential noise on images which could occur from different sources such as hardware or software. In this paper, we present a raw chicken portion identification system that uses a database of noisy and noise-free images to train a neural network to identify the different chicken portions. Experimental results demonstrate the robustness of the proposed identification system, and suggest that it can be effectively used with high accuracy in a poultry processing plant.