Exploring the Potential of A-ResNet in Person-Independent Face Recognition and Classification

Several studies have been published such as the LFW benchmark's upper bound for naive-deep face recognition has been studied by Zhou et al. [1]. They started by looking into how data distribution and size have an impact on system performance. They use a variety of cutting-edge approaches that have been developed in earlier literature to describe their findings when they have a sizable training dataset. They summarised their findings by stating that classification, feature extraction, and face detection are the three primary issues that need to be resolved in order to improve face recognition.. The data is biased and the false positive rate is relatively low. Iqbal et al. [2] have investigated face detection using angularly discriminative features and Deep Learning (DL). To reduce model errors, they have been suggested in classification strategies. On the LFW dataset, they ultimately obtained 99.77% accuracy. Balaban [3] has suggested cutting-edge DL and facial recognition. To benchmark these systems, the authors have emphasised the need for larger and more challenging public datasets. The joint identification for DL face representation was suggested by Sun et al. [4]. By using DL and both face recognition and verification signals, they demonstrate in this study that it is possible to do so successfully. On the LFW dataset, supervised face verification accuracy of 99% was achieved. Table 1 lists the accuracy that the aforementioned investigations were able to obtain.

Several studies have been published such as a computational framework for brain-inspired face recognition has been put out by Chowdhury et al. [5]. This work presents a novel idea for an ideal computational model of facial recognition software that incorporates both engineering counterparts of these cues from earlier studies and signals from the distributed face recognition mechanism of the brain. They discovered that accuracy decreased on average by 4%. In their study on face identification, Mao et al. [6]V employed a deep residual pooling network. They provide a complete learning architecture for recognising textures that integrates the CNN model's prior residual pooling layer for effective feature transfer. According to their claims, the dataset is randomly split into 60% for training and 40% for testing. Deep fair models for complex data labelling in graphs and explainable face recognition for Local Binary Pattern (LBP) have been developed by Franco et al. [7]. Their model's accuracy increased by 5%. In the future, they plan to extend their research to a wider variety of architectures and datasets, providing new information and guidelines on how to build more equitable models for challenging input data. An LBP face recognition survey system was proposed by Kortli et al. [8]. The strategies based on local, holistic (subspace) and hybrid characteristics are highlighted in this paper's summary of recent research on 2D and 3D face recognition systems. Additionally, they asserted that they have compared the processing speed, complexity, discrimination, and resilience of numerous approaches. Utilizing a super-wide regression network, Liu et al [9] have investigated and researched unsupervised cross-database facial expression identification. In this study, they provide a Special Super Wide Regression Network (SWiRN) model that serves as the regression parameter to connect the original feature space and label space.

A face image library called Labeled Faces in the Wild (LFW) was developed to study the issue of unrestricted face identification [10]. This database was created and is kept up to date by researchers at the University of Massachusetts, Amherst. 13,233 images of 5,749 people from the internet were recognised and centred using the Viola-Jones face detector. 1,680 of the people in the dataset had two or more different images. Four sets of LFW images and three different types of “aligned” images are included in the original database. The researchers found that for the majority of face verification techniques, deep-funnelled images performed better than alternative image formats. The dataset offered here is therefore the deep-funnelled form.

Since the 1970s, face recognition has been the subject of extensive research. To extract the faces from an input image that contains many faces, face detection is typically used by face recognition systems. A low-dimensional representation (or embedding) is produced and acquired after preprocessing each face. It is necessary to have a low-dimensional representation for effective classification. Face identification is challenging since faces are not solid objects and images might be taken from different perspectives. Face representations must be impervious to intrapersonal image fluctuations like those caused by age, expression, and style while yet being able to distinguish between interpersonal image variations between people. The preprocessed and enhanced input images are as follows:

They are scaled to fall between [0, 1].

The technique is broken down into two sections in this section: general methodology and the suggested model design.

The broad methodology primarily consists of two things:

1)

Face position

The authors begin by using face.evoLVe.PyTorch and MTCNN [11] for automatic face alignment. Figure 1 depicts the architecture of the deep cascaded multi-task framework that MTCNN proposes to improve ResNet's performance on face alignment by utilising their intrinsic correlation.

Figure 1.

However, the authors discover that even though MTCNN is quick, it occasionally makes mistakes and introduces erroneous data, such as in Figure 2, and these erroneous data will unquestionably have disastrous effects on model training. The authors then use the face-align tools from face.evoLVe to ultimately obtain accurate data. You may find this utility here [12]. This tool brings no dirty data but is around four times slower than MTCNN. However, the scientists are perplexed as to why MTCNN produces such incorrect results given that it is almost cutting edge. The face.evoLVe tool was created using the MTCNN. The authors evaluate a number of options, and the results demonstrate that when the default minim-windowsize is undefinable, MTCNN starts at 10×10 and has a propensity to obtain incorrect faces. Therefore, all results are positive once the authors set the minimum size at 40×40.

Figure 2.

The limited scale of the data the authors have makes it difficult to train a model without over-fitting. The authors believe that it is acceptable to fine-tune a model that has already been trained. The last levels must be created by the authors. The suggested model design primarily consists of two things:

1)

Pre-trained ResNet

The pre-trained weight that the authors download is a version of Google's FaceNet. The high-level model structure of FaceNet is depicted in Figure 3 [13]. They apply triple loss and ultimately achieve 0.997 accuracy at LWF.

Figure 3.

The first model, which was created to be improved upon by FaceNet, is tuned by the authors using Inception-ResNet [14]. Figure 4 depicts the Inception-ResNet architecture.

Figure 4.

2)

Altered-ResNet (A-ResNet)

The authors altered the final layers of the ResNet before testing which model performed best. According to the code snippet, the model has six final levels as shown in Figure 5.

Figure 5.

As a result, the authors wish to remove the layers after Conv2d, utilise some of their algorithms, and just update the final layers to include an additional 104 faces. This is because earlier levels contained the fundamental data necessary to recognise face traits and fundamental characteristics. In the modified model, the last linear, pooling, batchnorm, and sigmoid layers have been removed, leaving only a torch model. In order to leverage the features retrieved by Cov2d layers, the authors then construct a final layer's class with sample Flatten and Normalize layers. Figure 6 depicts the architecture in this manner. It can be called A-ResNet. The authors will train these two models and provide some details to determine the best in the following part.

Figure 6.

In deep learning, the optimizer is crucial, and different optimizers can perform completely differently. As is well known, “Adam” is a highly effective optimizer, but should authors also utilise it in their work? Figure 7 displays the outcomes of the authors' testing of RMS-prop, another theoretically sound optimizer, in Tensorborad-X.

Figure 7.

It demonstrates that the loss of the RMS optimizer actually decreases very quickly in the initial stages, and finally converges at a value of roughly 4.5. However, Figure 8 illustrates how much better the Adam optimizer performs with the identical epochs and batch sizes of 32 and 128.

Figure 8.

The findings on Inception-ResNet that the authors obtain after selecting various epoch and batch size combinations are shown in Table 2. More batch size typically results in improved performance, as shown in Table 2, although sometimes more epochs are required to minimise the loss. For example, 256 batch size performs worse than 128 batch size in 24 epochs before improving in 32 epochs. Finally, the ResNet achieves 82 true positives at 24 epochs, 128 batch size, and highest performance. The authors can quickly select a few combinations for A-ResNet using Table 2, and the outcomes are given in Table 3.

Fortunately, the A-ResNet outperforms ResNet at its peak performance of 24 epochs, 128 batch size, and 81 true positives. The authors are therefore pleased to declare that A-ResNet has won this combination with 10 more ture-positives. The least loss for ResNet during training is approximately 0.27, whereas the minimum loss for A-ResNet is approximately 3.8. This likely indicates that ResNet is constructed more intelligently in order to track and minimise the loss.

The outcomes of face recognition for a number of low-resolution input images are covered in this section. Table 6 displays the identification rates utilising our created database LFW and a rotating head around the camera. As image resolution improves, the identification rate rises. Additionally, a key element in determining recognition accuracy is the quantity of images in the database. Table 7 shows that the findings demonstrate that as the input images' pixel count increases, so does the recognition accuracy. Identification accuracy is strong even when the camera is surrounded by a moving head. This is because when the head is angled toward the camera, the cropped face image is aligned before being recognised, increasing recognition precision.

Even with the high number of epochs and steps in each epoch, the system performed with a testing set accuracy of 99.15% and a training set accuracy of 98.35%, proving that the model was not overfit (75 epochs of 276 steps). The A-ResNet's accuracy suggests that determining whether a face is wearing a mask is an easy problem to solve. The mask recognition model is not the most challenging aspect of the system, as has been discovered in earlier research on the subject. The real challenge is finding the locations of hidden faces in images. The authors classified the faces by using the A-ResNet and the Haar Cascade facial recognition system. The model worked well, according to the authors' manual examination of group images. Since the Haar Cascade approach required unique parameters for each image, this system is not automated, but it serves as a proof-of-concept for the model's ability to function with real-time input. Figures 9 (a and b) findings demonstrate that only actual faces are detected, and each face is correctly identified. Each classification is also accurate. Given that the model can classify each image in as little as 200 milliseconds and that the Haar Cascade technique can operate in real-time on a video stream, it is clear that once a facial detector is made autonomous, the model itself might be used to process real-time, on-the-fly data.

Figure 9.