Transfer learning based effective emotional face recognition using DCNN via cropping techniques

Facial Expression Recognition (FER) has grown in popularity as a result of the recent advancement and use of humancomputer interface technologies. Because the images can vary in brightness, backdrop, position, etc. it is challenging for current machine learning and deep learning models to identify facial expression. If the database is small, it doesn't operate well. Feature extraction is crucial for FER, and if the derived characteristics can be separated, even a straightforward approach can help tremendously. Deep learning techniques and automated feature extraction, allow some irrelevant features to conflict with important features. In this paper, we deal with limited data and simply extract useful features from images. To make data more numerous and allow for the extraction of just important facial features, we suggest innovative face cropping, rotation, and simplification procedures and advocate using the Transfer Learning technique to construct DCNN for building a very accurate FER system. By replacing the dense top layer(s) with FER, a pretrained DCNN model is adopted, and the model is then modified with facial expression data. The training of the dense layer(s) is followed by adjusting each of the pre-trained DCNN blocks in turn. This new pipeline technique has gradually increased the accuracy of FER to a higher degree. On the CK+ and JAFFE datasets, experiments were run to assess the suggested methodology. For 7-class studies on the CK+ and JAFFE databases, high average accuracy in recognition of 99.49% and 98.58% were acquired.