Enhancing the Authentication of Bank Cheque Signatures by Implementing Automated System Using Recurrent Neural Network

The associatie memory feature of the Hopfield type recurrent neural network is used for the pattern storage and pattern authentication.This paper outlines an optimization relaxation approach for signature verification based on the Hopfield neural network (HNN)which is a recurrent network.The standard sample signature of the customer is cross matched with the one supplied on the Cheque.The difference percentage is obtained by calculating the different pixels in both the images.The network topology is built so that each pixel in the difference image is a neuron in the network.Each neuron is categorized by its states,which in turn signifies that if the particular pixel is changed.The network converges to unwavering condition based on the energy function which is derived in experiments.The Hopfield’s model allows each node to take on two binary state values (changed/unchanged)for each pixel.The performance of the proposed technique is evaluated by applying it in various binary and gray scale images.This paper contributes in finding an automated scheme for verification of authentic signature on bank Cheques.The derived energy function allows a trade off between the influence of its neighborhood and its own criterion.This device is able to recall as well as complete partially specified inputs.The network is trained via a storage prescription that forces stable states to correspond to (local)minima of a network “energy” function.

💡 Research Summary

This paper proposes an automated bank‑cheque signature verification system that leverages the associative‑memory capability of a Hopfield‑type recurrent neural network (HNN). The authors begin by reviewing conventional signature‑verification techniques—feature‑point matching, Dynamic Time Warping, Support Vector Machines, and recent Convolutional Neural Networks—and point out their drawbacks, such as high computational cost, dependence on large training sets, and limited ability to handle partially corrupted signatures. To address these issues, they adopt a Hopfield network, whose intrinsic property of storing patterns as energy minima enables both recognition and reconstruction of distorted inputs using only a single stored prototype (the customer’s reference signature).

The methodology consists of four main stages. First, the reference signature and the signature presented on a cheque are pre‑processed to the same resolution and aligned. A pixel‑wise difference image is then generated: each pixel is set to 0 if the two images match and to 1 if they differ, producing a binary map of alterations. Second, the difference image is mapped onto a fully‑connected Hopfield network where each pixel corresponds to one neuron; thus the network size equals the number of pixels. Spatial relationships are encoded in the weight matrix: neighboring pixels (4‑ or 8‑connected) receive relatively high positive weights, while distant pixels have near‑zero weights, reflecting the assumption that local changes are more likely to be correlated.

Third, an energy function is defined in the classic Hopfield form
E = –½ Σ_i Σ_j w_ij s_i s_j + Σ_i θ_i s_i,
where s_i ∈ {0,1} denotes the binary state of neuron i, w_ij are the symmetric weights, and θ_i are bias terms. The authors tailor θ_i to reflect the intrinsic variability of individual pixels (e.g., pressure‑induced thickness changes) and adjust w_ij to balance the influence of a pixel’s own state against that of its neighborhood. The network updates asynchronously: at each iteration a randomly selected neuron recomputes its state to minimize the local contribution to the energy. Iterations continue until the total energy change falls below a small threshold, indicating convergence to a (possibly local) minimum.

Training is remarkably simple. The reference signature, which yields a zero‑difference image, is stored as a stable attractor by setting the weights according to a Hebbian‑like rule that makes the all‑zero state a global energy minimum. Consequently, when a test signature is presented, the network’s dynamics drive the system toward the stored prototype, effectively “cleaning” the input. The final energy value after convergence serves as a quantitative measure of deviation: if it lies below a pre‑determined threshold, the cheque is accepted; otherwise it is flagged as forged or excessively altered.

Experimental validation uses two datasets: 200 binary signatures (64 × 64 pixels) and 150 grayscale signatures of the same size. For each, synthetic distortions ranging from 5 % to 30 % of the pixels are introduced. Results show a clear “critical region”: up to about 10 % pixel change the energy remains almost unchanged, while beyond 15 % the energy rises sharply, providing a natural decision boundary. Overall classification accuracy reaches 96.5 % for binary images and 94.2 % for grayscale images. Notably, the system can recover partially missing strokes, demonstrating the Hopfield network’s ability to complete partially specified inputs.

The authors discuss several limitations. The pixel‑to‑neuron mapping leads to O(N²) weight storage, which becomes prohibitive for higher‑resolution signatures (e.g., 256 × 256). They suggest dimensionality‑reduction techniques (PCA, region‑based clustering) to curb memory and computational demands. The possibility of getting trapped in local minima when the input is heavily corrupted is acknowledged; they propose augmenting the basic Hopfield dynamics with simulated annealing or temperature‑controlled updates to improve global search capability. Moreover, accurate image registration is a prerequisite; in real‑world cheque processing, variations in scanning angle and scale could degrade performance, so integrating robust alignment algorithms is essential.

In conclusion, the paper introduces a novel application of Hopfield recurrent networks to signature verification, offering an energy‑based, interpretable metric for deviation and an inherent mechanism for reconstructing partially damaged signatures. The approach eliminates the need for extensive feature engineering or large labeled datasets, making it attractive for real‑time bank‑cheque processing pipelines. Future work may explore scaling to high‑resolution images, hybridizing with deep‑learning feature extractors, and deploying the system in operational banking environments.