Deep-Learning Approach for Efficient Eye-blink Detection with Hybrid Optimization Concept

Abstract

In this research work, a novel eye-blink detection model is developed. The proposed eye blink detection model is modeled by following seven major phases: (a) video-to-frame conversion, (b) Pre-processing, (c) face detection, (d) eye region localization, (e) eye landmark detection and eye status detection, (f) eye blink detection and (f) Eye blink Classification. Initially, from the collected raw video sequence (input), each individual frames are extracted in the video -to-frame conversion phase. Then, each of the frames is subjected to pre-processing phase, where the quality of the image in the frames is improved using proposed Kernel median filtering (KMF) approach. In the face detection phase, the Viola-Jones Model has been utilized. Then, from the detected faces, the eye region is localization within the proposed eye region localization phase. The proposed Eye region localization phase encapsulates two major phases: Feature extraction and landmark detection. The features like improved active shape models (I-ASMs), Local Binary pattern are extracted from the detected facial images. Then, the eye region is localization by using a new optimized Convolution neural network framework. This optimized CNN framework is trained with the extracted features (I-ASM and LBP). Moreover, to enhance the classification accuracy of eye localization, the weight of CNN is fine-tuned using a new Seagull Optimization with Enhanced Exploration (SOEE), which is the improved version of standard Seagull Optimization Algorithm (SOA). The outcome from optimized CNN framework is providing the exact location of the eye region. Once the eye region is detected, it is essential to detect the status of the eye (whether open or close). The status of the eye is detected by computing the eye aspect ratio (EAR). Then, the identified eye blinks are classified based on the computed correlation coefficient as long and short blinks. Finally, a comparative evaluation has been accomplished to validate the projected model.