期刊：Azerbaijan journal of high performance computing
[Azerbaijan State Oil and Industry University]
卷期号：4 (1): 60-90
Human gender recognition is one the most challenging task in computer vision, especially in pedestrians, due to so much variation in human poses, video acquisition, illumination, occlusion, and human clothes, etc. In this article, we have considered gender recognition which is very important to be considered in video surveillance. To make the system automated to recognize the gender, we have provided a novel technique based on the extraction of features through different methodologies. Our technique consists of 4 steps a) preprocessing, b) feature extraction, c) feature fusion, d) classification. The exciting area is separated in the first step, which is the full body from the images. After that, images are divided into two halves on the ratio of 2:3 to acquire sets of upper body and lower body. In the second step, three handcrafted feature extractors, HOG, Gabor, and granulometry, extract the feature vectors using different score values. These feature vectors are fused to create one strong feature vector on which results are evaluated. Experiments are performed on full-body datasets to make the best configuration of features. The features are extracted through different feature extractors in different numbers to generate their feature vectors. Those features are fused to create a strong feature vector. This feature vector is then utilized for classification. For classification, SVM and KNN classifiers are used. Results are evaluated on five performance measures: Accuracy, Precision, Sensitivity, Specificity, and Area under the curve. The best results that have been acquired are on the upper body, which is 88.7% accuracy and 0.96 AUC. The results are compared with the existing methodologies, and hence it is concluded that the proposed method has significantly achieved higher results.
Strongly Powered by AbleSci AI