Back Propagation Neural Network and Chicken Swarm Optimization for Yoruba Indigenous Food Recognition

Abstract

Abstract

Back Propagation Neural Networks (BPNNs) are widely used to model complex systems in a steady state.

However, BPNNs have a slow convergence rate. Several meta-heuristic algorithms have been used to speed up its convergence. Though, the modifications increased the number of parameters which affect the convergence rate, but more work needs to be done on BPNN in order to improve the performance of BPNN. This work introduced Chicken Swarm Optimization (CSO) technique to improve the performance of BPNNs in order to recognize Yoruba indigenous food. 1251 varieties of Yoruba indigenous foods such as Amala ogede, Iyan gbere, Puguru, Akara, Dele, Ekuru, Monu, Aseke, Abari, Sapala, and Egbo were prepared. The prepared foodswere captured using a digital camera and were digitalized. The digitalized foods were subjected to preprocessing using Image resizing, Morphological, Edge scaling, Histogram, normalization and Sobel edge. All images were subjected to feature extraction using Local Binary Pattern (LBP) and the feature vectors gotten were optimized with Chicken Swarm Optimization (CSO) algorithm. The result of the optimized parameters were classified using Back Propagation Neural Network (BPNN). The recognition accuracy using BPNN yields 82.22 %, 87.59 %, 83.67 %, and 85.34 % for the four categories of food respectively, whereas BPNN-CSO yields 85.34 %, 91.90 %, 91.02 %, and 90.23 %, respectively. Sensitivity using BPNN yields 87.96%, 92.22%, 90.22%, 90.13% while BPNN-CSO delivers 90.13%, 97.78%, 97.74%, 96.71% respectively using recognition time and sensitivity standard term that are included in the result table. It was observed that, the recognition accuracy, sensitivity, specificity, and false positive ratio values of BPNN-CSO gives improved performance result compared to BPNN.