Hybrid CNN-Transformer for Dynamic Indian Sign Language Recognition with Non-Manual Gesture Analysis

Sign language is a highly dense, visually based interaction, whose effectiveness is predetermined by a highly organized system of hand gestures in space and the concomitant use of musculature of the face, eye movements, articulatory gestures, and other non-manual signatures, which gives the speech of sign language an artistic shift of power [1]. There have arisen all over the world three hundred or so different sign systems, each of them diversifying iteratively in the specificity of the geographical region, the texture of the culture, and the subtlety of the language. Similar to the idea that is presented in spoken tongues of lingering dialects and accent, sign languages display substantial deviation when it comes to gesture construction, sequential correctness, idiomatic expression, and non-manual semiotic adornment. Especially, it is worth mentioning that Indian Sign Language (ISL) takes a prominent place in the South Asian linguistic landscape, which is the symbol of diversity and epistemic richness that defines the spectrum of sign languages [2]. Whereas sign languages can be coarsely categorized into the group of static gestures, where the hand shapes are created by one or both hands, and the group of dynamic gestures, where the temporal movement of the hands and the variations of the expression are added, the overwhelming part of the extant research has either overemphasized the former classification or oversimplified the current systems of gestures. This chauvinism poses a great gap in full identification of the broader spectrum of sign language that embraces facial expression and other more elaborate body languages [3]. Furthermore, the dominating paradigms are based on the recognition of transient manual movements or on those that are written in a speech form, thus overlooking the complete repertoire of sign-language communication and limiting the effectiveness of human-computer interaction systems that have been developed to correspond with the deaf community. There is an urgent need for powerful sign language recognition systems that go beyond the rigid gestures and written texts. We aim to develop a state-of-the-art solution that could understand dynamic sign language sufficiently well to include subtle facial expressions, among others, with the employment of state-of-the-art deep-learning algorithms. Precisely, this paper aims to:

1. Propose and combine a Vision Transformer (ViT) to enhance the Convolutional Neural Network (CNN) ability to identify sign language in video footage.

2. Use a sample of naturally deaf and trained Sign Language speakers so as to incubate the intricacy of sign languages as they are truly utilized.

3. A gap in the existing work of research will be resolved by focusing on dynamic, continuous signs with non-manual attributes, thus going beyond the static signs and crude gestures that have been prevalent in the literature.

The contribution to this paper is introducing a video-based sign-language recognition model which uses Vision Transformer to process dynamic signs, facial expressions of nuanced appearances and other non-manual societal readings. In this way, we get the field to a stage where it is not constrained by its own limitations, where more inclusive and effective solutions to human-computer interaction are possible that realize the richness and the complexity of the sign-language interaction. After this introduction, Section 1 is the review of related work in sign-language recognition. Section-2 explains the suggested methodology and elaborates on how it will address the challenges mentioned. Section 3 describes the architecture of Vision Transformer and its video recognition application in detail. Section 4 describes and discusses the results of the experiment. Finally, Section 5 wraps up this research providing a summary of the main findings and future research avenues.

Conventional CNN-based models which have historically been used in image and video tasks have good local feature extraction properties. Nevertheless, they in many cases use sequential feature summation (e.g., through RNNs or 3D convolutions) to do so, which can be costly and not always optimal to establish long-range correlations in videos. Whereas purely transformer-based schemes (ViT) can capture both the global and temporal context and find local differences, they can ignore small variations in local context that can distinguish similar gestures.

Through the incorporation of CNN layers into the ViT structure we have been able to maintain the local pattern recognition capabilities of CNNs whilst still exploiting the ViT capability to capture multi-frame complex temporal and contextual interactions. This synergy does enhance recognition and is especially accurate with complex ISL gestures which are based on both fine-grained hand configurations and complex temporal patterns.

CNN-Only Models:

CNN-based methods in pure form might be too restricted in terms of ability to realize the temporal global context unless other components, including RNNs or 3D convolutions, are added to them. In our initial experiments, we found that, because of the rich spatial feature retrieval capabilities of 3D CNNs, their performance in interpreting long sequences of behaviors was lower than that of the ViT-based counterparts.

RNN or LSTM-Based Models:

Though RNNs or LSTMs are capable of modeling temporal sequences, they can be more susceptible to such problems as vanishing gradients across long videos and even less efficient than attention-based models. Initial experiments with CNN-LSTM hybrids provided decent performance at the price of increased training time and reduced modeling capabilities of complicated time-dependent dependencies.

Pure ViT Models:

Pure ViT models are superior in modeling long-range dependencies. Nevertheless, they do not have any local feature extraction element and hence do not pick up finer details such as minute movements of the fingers or micro-expressions on faces. Experiments on our part showed that the general fine-grained recognition accuracy was enhanced with the addition of CNN layers, especially in harsh backgrounds.

Altogether, the hybrid ViT-CNN solution provides a middle ground solution based on global time modeling (transformers) and local space feature extraction (CNNs). Empirical studies demonstrate that this hybrid architecture outperforms purely CNN-based or purely ViT-based models and are comparable to CNN-RNN hybrids in accuracy, efficiency and generalization to various ISL gestures.

3.1 Flow of the study

The current research aims to develop an effective ISL recognition application that could decode dynamic hand gestures and facial expressions on the basis of video streams correctly. Where a large number of research focuses on gestural frames or static gestures, this research focuses on continuous gestural signs, including small video recordings of 1 to 3 seconds. In this direction, we choose a hybrid neural network combining the capabilities of global attention of a ViT with the potential of a local feature-extraction of a CNN. The whole processing chain is described in Figure 1 and summarized in the following parts:

Figure 1. Block diagram

3.1.1 Video capture

At the first stage, we obtain a rich set of video recordings of deaf subjects instructing a repertoire of predetermined ISL gestures in a range of lighting scenarios and background settings, which is used to strengthen the later system against environmental covariate influences.

3.1.2 Frame extraction

The videos that are captured are then divided into discrete frames; these temporal snapshots are the two elements of the manual component, i. e., the hand kinematics, and the non-manual component, i. e., the facial musculature, both of which cannot be done without reading the signs in any manner.

3.1.3. Frame counting and padding

We also come up with the frame tally of both records and the temporal maximum over the corpus; those videos that are shorter are electronically padded to that temporal maximum which normalizes the input dimension of all further processing phases.

3.1.4 Landmark detection

A Boolean landmark-detection processes on each frame provide a list of salient features, most likely to occur around the hands and face of the signer, that provide a structured, fine-grained image of the gestures and subtle facial expressions. A similar approach is seen in the work of Jo et al. [28] for enhancing gestural interaction used in virtual and augmented reality with Media-Pipe based gesture recognition interface.

3.1.5 Position encoding

The network takes spatial context through positional encodings to the retrieved landmarks; this process embeds inter-point spatial relations and relative positions of anatomy parts into the model and provides it with a better understanding of gestural structure.

3.1.6 Mask computation

A saliency mask that gives more weight to the hand, face, and other regions of interest is synthesized by us and effectively reduces background clutter, sensor noise, and focal capacity of the model, focusing attention on the informative regions of the spatial map.

Development: ViT architecture: ViT is an architecture using artificial intelligence (AI) to recognize images as objects and extract information from them.

Motivation: ViT architecture: ViT is a type of architecture based on AI that identifies objects in images and derives information about objects contained within the image. Instead of defining images, or sequentially ordered frames, as a set of convolutional filters, Vision Transformers conceptualize images as a set of tokenized patches and use multi-head self-attention systems to encode long-range correlations and contextual interactions. We then subdivide the extracted frames (or their spatial encodings, e.g., landmark-based ones) into small patches, and in this manner, a linear projection of the patch into a high-dimensional embedding space and positional embedding is generated.

3.1.7 Feature fusion

The unified approach of combining CNN with Transformer follows a sequential feature encoding and fusion mechanism. For each input video sequence, individual frames are first processed by a backbone of CNN, which extracts fine-grained spatial features corresponding to hand shape, finger articulation as well as non-manual cues like facial expressions and head orientation. The embeddings generated by CNN are then temporally arranged and sourced as token representations to the vision transformer where self-attention mechanisms model long-range sequential dependancies across frames. The transformer output is lastly fused with the CNN features by concatenating before classification.

Mathematical Formulation of the model:

Let

$X=\{x 1, x 2, x 3, \ldots, x T\}$                 (1)

be the video with T frames.

CNN feature extraction would be

$f_t={CNN}\left(x_t\right) ;$ where $f_t \in \mathbb{R}^d$                    (2)

Stacked spatial matrix is represented by:

$F=\left[f_1, f_2, f_3, \ldots, f_T\right]$                        (3)

Position encoding is depicted by

$Z=F+P$                   (4)

where, P is positional encoding matrix and Z is position aware feature representation. The Transformation attention spaces query (Q), key (K) and value (V) are represented as

$Q=Z W_Q, K=Z W_K\ and\ V=Z W_V$                 (5)

The transformer output, fusion representation, and classification equation are as follows:

$T_f= Transformer(Z)$               (6)

$F_{ {fusion }}=\left[{Pool}(F) ; {Pool}\left(T_f\right)\right]$                          (7)

$y={Softmax}\left(W F_{{fusion }}+b\right)$                         (8)

where, Pool represents global maxpooling and [;] indicated concatination of feature vectors.

The method of fusion is as depicted in Figure 2. This fusion design enables the model to preserve fine – graned spation discrimination through CNN feature extraction while simultaneously exploiting the Transformer’s ability to capture temporal features. This synergy is particularly beneficial for dynamic ISL gesture recognition.

Figure 2. CNN- ViT fusion mechanism

3.1.8 ViT and adapting to the Desired Task:

(a) Temporal Patching: Each frame or the set of landmarks that accompanies the frame is treated as a separate token, thus the transformer can focus on the spatial axis and simultaneously the time axis.

(b) Temporal Positional Embedding: It makes the network self-temporal: the network is given the ability to represent its dynamic properties of motion and gesture development over time by encoding order in vectors indicating frame chronology.

The application of CNN Elements and ViT Behavior:

Although ViT is highly successful in observing the dependencies of the world, its peculiarity of using pure attention might not always adequately reflect local, subtle forms (such as fine grasping configurations or subtle facial micro-expressions). In order to address this weakness, we utilize CNN modules as part of the ViT pipeline. In everything hybrid, raw frames pass through a lightweight CNN that isolates salient primitives in space: edges and textures, contours, etc., before being subjected to the transformer. CNN feature maps such that result are in turn fused with the ViT embedding, allowing the transformer layers to have access to the enriched inputs that combine the benefits of global context framing with the benefits of distilled local detail.

3.1.9 SoftMax layer

The unprocessed output of the hybrid ViT-CNN processing is then fed into a SoftMax classifier to recede to an emergent representation, enabling categorical probabilities of every gesture classes repertoire of ISL to be produced, hence allowing decisive identification of signs with each input sequence. To identify the relationship, the model undertook and its strength, the tests involve training, validating, and testing the model.

The ViT-CNN architecture was trained and optimized on the curated dataset of ISL and the following hyper-parameters and regularization options were used:

o Learning Rate: Learning rate at the beginning will be 1e -4 and as plateau in the validation accuracy is reached, it will be decreased by a factor of 0.1.

o Number of Epochs: 50 -100, based on convergence patterns.

o Batch Size: 8-16, which was selected according to the available memory of the GPUs and stability of the training.

o Regularization Techniques: To address overfitting, dropout layers (dropout rate of 0.3–0.5) are included not only in the CNN layers but also in the transformer ones. Also, early stopping and data augmentation are applied (e.g. random cropping, limited rotations) to guarantee improved generalization.

3.1.10 Model scoring

After the training process, a detailed assessment of the withheld test split is performed where we calculate conventional performance measures such as accuracy, precision, recall and the F1-score to objectively assess the efficacy of the system in both controlled laboratory and real world background circumstances. This test confirms that the model can be dependable in distinguishing between twenty-two different classes of ISL gestures.

3.1.11 Inference

When its validation with sufficient accuracy is achieved, the hybrid ViT-CNN pronounces a real-time inference feature. Once a new video stream is consumed, the system will run landmark extraction, positional encoding, and mask generation sequentially and run the content through the hybrid network, producing an estimation of a gesture and a confidence estimate in the end.

This paper is dedicated to strict analysis and subtle understanding of naturally pre-established ISL frameworks, and the major aim to outline the gap in communication between the disabled population and the rest of the community.

The study combines ViT modules with traditional CNN models by building an indigenous ISL dataset by recording participants who are deaf and therefore enhancing the recognition accuracy.

The sensitive application of encoder transformer avoids the use of complicated data preprocessing, and the discriminating addition of the ViT highlights the strength and performance measure of the model.

Model Summary:

Table 1 provides a concise overview of the model architecture and the cumulative trainable parameters.

Table 1. Model summary

Training and Validation Performance:

During the course of training, significant success is found. Training accuracy reaches amazing 100‛, which shows the ability of the model to generalize the training data. The training accuracy is 95, and the recall is 92 with the resulting F1-score of 0.95 in 1,393 epochs. These measures highlight the capability of the model to pick the positive instances correctly and have a balance between the precision and recall.

During validation stage, the model maintains a healthy performance. The agreement of validation stabilizes to 88.60 per cent, and the precision is 87 and the recall is 86. Validation F1-score: 0.89, realized in 2,987 epochs. These statistics show that the model is a good generalization, which can still maintain a good performance when it is applied to data that is not seen.

Construction: testing Performance and Comparative Analysis:

The model when rigorously tested on another set of 55 videos gives a testing accuracy of 82.14. Accuracy is 81.89, and recall is 81.36 with an F1-score of 0.81 in 3,000 epochs. Such findings support the effectiveness of the model in identifying ISL gestures in real world situations.

These performance metrics point out the consistency of proficiency of the model in training, validation and testing phases.

Table 2. Model performance

Class -Level Performance and Error Analysis:

To gain more insights, the performance of the model was analyzed according to the classes, presented in Table 2. And Figure 5 can be seen as the confusion matrix. The results evidently show strong validation of the system’s suitability to correctly classify signs like Namaste, Hello, Danger, Help Me, and I am Hungry. This proves the system’s ability to perform practical ISL translation as the signs included are for: Greeting, Emergency Communication and conveying basic needs. Classification summary indicated that the features which are visually and dynamically distinct lead to near- perfect separability. The most contributing features to classification are Hand Shape, Orientation, Motion Pattern and Semantic uniqueness of the gestures. Although the majority of classes have high F1-scores, some of the classes such as “Ten” have relatively lower F1. In order to gain more insight into these discrepancies, a confusion matrix was obtained, as shown in Figure 5. The confusion chart demonstrates that gestures with similar hand shapes or orientations were wrongly classified quite often. The confusion matrix has strong diagonal dominance. An example is that the model was more likely to confuse the signs that have similar fingers arrangement or indicate slight rotations in their hands.

Figure 5. Confusion matrix

Some of the difficult classes include, but are not limited to, classes: Ten and Nine which exhibit a significant level of inter-class confusion as a result of the fact that these two classes are similar in their hand shape and position. Equally, those signs that require rapid changes or faint body expressions were also mistaken. The model may not be able to differentiate between gestures that are differentiated by slight variance in finger positioning or slight movement of the wrist and such therefore will be misclassified. The classification results can be grouped in 3 distinct categories. Group A: Perfectly Classified Classes, Group B: Moderately Strong Classes, Group C: Weak Classes as seen in Table 3. We can see that the gestures where distinctive sand shapes and motions are dominant have been precisely classified with a high F1-score. The gestures where partial feature overlap is possible within the neighboring frames are moderately low in F1-score. The most misclassified signs are too similar in nature. The hand posture, orientation and visual similarity between the signs One, Ten and Hundred are too close. The analysis indicated that the error pattern is not random, however semantically more meaningful. Visual Resemblances of Gestures here stand out as a possible cause of misclassification. Little movements in bending fingers or the position of thumbs can be very hard to detect by the model. The results demonstrate that the proposed framework achieved excellent recognition for emergency, basic need and other conversational gestures, while giving challenges in some numeric sign gestures with similar visual patterns due to high inter-class similarity and subtle articulation differences.

Table 3. Grouping of classes as per performance of the recognition

Even specific refinements (like a more careful data augmentation, better landmark detection, or using other cues, like depth or skeleton data) can be made by looking at cases of misclassification and understanding their causes.

By examining misclassifications as depicted in Table 4 and understanding their underlying causes, targeted improvements—such as more elaborate data augmentation, refined landmark detection, or incorporating additional cues (like depth or skeleton data)—can be implemented.

Table 4. Class wise performance

Consequences of the Real-World Usage:

The noted matter of confusion has an important implication for the usage of the sign recognition system in the life environment. This is a must in the daily communicative interactions of a system where slight differences in gestures must be dealt with high accuracy. In other settings, including educational institutions, clinical facilities, or customer service touchpoints, misclassification of a particular gesture could trigger the occurrence of an expensive misunderstanding. This may lead to a necessity on the part of practitioners to embrace more vivid signing conventions, or the training corpus may be expanded by engineers to a wider cohort of signer heterogeneity and ambient environmental situations.

Future work could be done by increasing the size of the corpus to include a wider range of difficult gestures examples.

Adding multi-modal sensors of sensory data (depth sensors or skeletal tracking) to provide a more detailed context system.

Optimizing the hybrid ViT-CNN like in a cross-head, i.e., adding to the model, to reinforce its local element extraction capabilities, especially on gestures of nearly the same shape.

Implementation of domain adaptation measures so that there is robustness to the context environment interactions and variations in the background.