Enhancing Object Detection Accuracy Using a Hybrid CNN–Transformer Architecture for Complex Visual Environments

The detection of objects has been one of the core systems of the contemporary computer-vision applications to include autonomous driving, surveillance, smart cities, and medical imaging and human-robot interaction. But the current generation CNN-based detentions have difficulties in capturing long-range contextual information with Transformer-based models working well in capturing global dependencies but because it incurs a high cost of computation and tends to weak local inductive biases. To overcome these shortcomings, this paper introduces a Hybrid CNN-Transformer architecture that would be used to improve the object detection capabilities on complicated visual scenes. The model combines a CNN backbone, which consists of fine-grained spatial feature extraction with a Transformer encoder which can extract global contextual relationships implicit in the image. The feature fusion module based on multi-scale features and the attention-guided detection head also enhance the capabilities of the model to identify small, occlusives, and frontal packed objects. As experimental tests on the COCO and Pascal VOC datasets show, the given model has a higher accuracy rate than those of the state-of-the-art detectors, the difference with them in terms of metrics (mean average precision) is +3-5 percent of mAP, and the small-object detection is more advanced. Although the model includes Transformer elements, its inference rate is competitive (24 FPS), which is why it can be applied to real-time settings and resource-limited scenarios. The findings validate the fact that a convolutional-Transformer-based combination is an effective and effective method of advancing object detection systems in the next generation.