Abstract
These days, the application of image processing in computer vision is becoming more crucial. Some situations necessitate a solution based on computer vision and growing deep learning. One method continuously developed in deep learning is the Convolutional Neural Network, with MobileNet, EfficientNet, VGG16, and others being widely used architectures. Using the CNN architecture, the dataset consists primarily of images; the more datasets there are, the more image storage space will be required. Compression via the discrete cosine transform technique is a method to address this issue. We implement the DCT compression method in the present research to get around the system's limited storage space. Using DCT, we also compare compressed and uncompressed images. All users who had been trained with each test 5 times for a total of 150 tests were given the test. Based on testing findings, the size reduction rate for compressed and uncompressed images is measured at 25%. The case study presented is face recognition, and the training results indicate that the accuracy of compressed images using the DCT approach ranges from 91.33% to 100%. Still, the accuracy of uncompressed facial images ranges from 98.17% to 100%. In addition, the accuracy of the proposed CNN architecture has increased to 87.43%, while the accuracy of MobileNet has increased by 16.75%. The accuracy of EfficientNetB1 with noisy-student weights is measured at 74.91%, and the accuracy of EfficientNetB1 with imageNet weights can reach 100%. Facial biometric authentication using a deep learning algorithm and DCT-compressed images was successfully accomplished with an accuracy value of 95.33% and an error value of 4.67%.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have