PERFORMANCE COMPARISON OF FASTER R-CONVOLUTIONAL NEURAL NETWORK (CNN) AND EFFICIENTNET FOR TRAIN DETECTION UNDER DIVERSE LIGHTING AND IMAGE QUALITY CONDITIONS

Object detection using computer vision has seen rapid advancements, especially with the advent of deep learning architectures such as Faster R-CNN and EfficientNet. This study compares the performance of the two models in detecting trains in various lighting conditions and noise disturbances. The dataset used consisted of 4500 train images which were divided into 70% for training, 20% for validation, and 10% for testing, reflecting real-world conditions. The evaluation was carried out using the Intersection over Union (IoU), Average Precision (AP), and Average Recall (AR) metrics. The results show that Faster R-CNN consistently excels in terms of detection accuracy, especially in less than ideal lighting conditions and under rain noise interference. In sufficient lighting conditions, Faster R-CNN showed a slightly superior AP score with a score of 0.844. As the lighting decreased, the difference between the two models became more pronounced, with Faster R-CNN recording an AP value of 0.810. In conditions with rain noise interference, the object detection performance of both models decreased more significantly, but the Faster R-CNN still excelled with an AP value of 0.798. Although EfficientNet is more efficient in terms of training speed, with a time of 5 hours and 37 minutes, and a smaller model size, Faster R-CNN shows higher reliability in complex environmental situations. This research provides important insights for the development of reliable and efficient train detection systems, taking into account the trade-off between resource efficiency and detection accuracy.