Muthanna Journal of Engineering and Technology
Volume (14), Issue (3), Year (2026), Pages (106-117)
DOI:10.52113/3/eng/mjet/2026-14-03-/106-117
Research Article By:
Barakat Saad Ibrahim and Tabarek Alwan Tuib
Corresponding author E-mail: tabarik.alwan@mu.edu.iq
ABSTRACT
Fine-grained car classification based on car type and model is crucial for intelligent transportation systems, traffic surveillance, smart parking and automated vehicle monitoring. However, the accurate classification still remains challenging because many car models have similar attributes in visual structure and scenes of the image samples may vary in viewpoint, illumination, size or even background information or occlusion. To correct the issues that could be remedied, .This paper proposed a PSO-optimized BiLSTM image-to-sequence system based on image (car) appearance for fine-grained car type and model classification. Developed through three components – ordered generation of image-patches sequence, bidirectional recurrent feature learning and Particle Swarm Optimization (PSO) hyperparameter optimization combined in a single classification pipeline. Conventional CNN classifiers typically focus on learning local spatial filters than model the spatial order, but the proposed approach represents each resized RGB vehicle image as a sequence of non-overlapping visual patches, enabling the BiLSTM to model the spatial order of the discriminative vehicle parts like the grille, headlights, roofline, wheels, and body contour. The model uses PSO to select the learning rate, dropout rate, number of BiLSTM units, number of recursive depths, dense-layer size, and batch size, providing less manual tuning of hyperparameters. The model was tested for performance on a hyper-cosine balanced 7 class vehicle image dataset which includes Hyundai Creta, Toyota Innova, Mahindra Scorpio, Audi, Swift, BMW and Mercedes Benz class. The accuracy of the proposed PSO-BiLSTM is 96.4%, precision 95.9%, recall 95.6%, and F1-score 95.7% compared to CPU baseline models CNN, RNN, BiLSTM, PSO-RNN, ResNet-50, MobileNetV2, EfficientNet-B0, ViT-B/16, and attention-based BiLSTM. The results found that it is possible to increase the discriminative power of similar vehicle categories for a bidirectional sequence learning task, while preserving accurate validation performance, through a swarm-based hyperparameter optimization.
Keywords: Fine-grained vehicle classification; car model detection ; image-to-sequence learning; BiLSTM; Particle Swarm Optimization; PSO; intelligent transportation systems.