PSO-Optimized BiLSTM Framework for Fine-Grained Car Type and Model Classification Using Image-to-Sequence Learning

الملخص

Fine-grained car type and model classification is an important task in intelligent transportation systems, traffic surveillance, smart parking, and automated vehicle monitoring. But the accurate classification is still difficult due to the fact that many car models may have similar visual structure, and the image samples may change their viewpoints, illuminations, scales and even background and occlusion. To address these issues, this study  proposes a Particle Swarm Optimization (PSO) optimized Bidirectional Long Short-Term Memory (BiLSTM) which is a framework of image-to-sequence learning for car type and model classification. In the proposed approach, each vehicle image is resized, normalized, and augmented to improve the model's resilience and data quality. The image is then subdivided into ordered visual patches, each of which is converted to a sequential feature vector and input into the BiLSTM network. This allows the model to express long-range spatial relationships between key vehicle features like headlights, grill construction, roof form, wheel design and body curves. The main hyperparameters of the BiLSTM model such as the dropout and learning rates, recurrent units, batch size, dense-layer size, and recurrent depth are automatically optimized using the PSO technique. The confusion matrix, F1-score, recall, accuracy, and precision analysis were utilized to assess the suggested PSO-BiLSTM model on a seven-class vehicle image dataset. The According to experimental findings, the suggested model outperforms the baseline CNN, standard RNN, BiLSTM and PSO-RNN model 96.4% accuracy, 95.9% precision, 95.6% recall, and 95.7% F1-score. These results show that the fusion of image-to-sequence learning, bidirectional recurrent modeling, and PSO-based hyperparameter optimization are conducive to improving classification performance and avoiding overfitting in fine-grained vehicle detection