An End-to-End Scratch-Trained Object Detection Framework with Local Attention and Multi-Scale Feature Fusion

A novel end-to-end object detection architecture is put forward to tackle the inherent limitations of existing models in scalability and generalization. Unlike state-of-the-art detectors that rely on networks pretrained on massive datasets such as ImageNet and COCO, this approach avoids the biases and constraints imposed by pretraining.through systematic architectural optimization, it achieves dual improvements in detection accuracy and extensibility, enabling continuous optimization of both speed and precision. The key innovation, a regional attention mechanism integrated into the architecture, enhances the representation of salient features and suppresses redundant information by dynamically allocating weights, and as a plug-and-play module, It can be smoothly incorporated into any network architecture to directly enhance performance. Experimental findings show that the model surpasses traditional baselines trained from scratch in real-time detection speed. Without employing pretrained weights, it achieves a mean Average Precision (mAP) of 71.2% on the VOC2007 dataset and 71.5% on the VOC2012 dataset. validating the effectiveness of the zero-pretraining paradigm and laying a theoretical foundation for the construction of adaptive and evolvable self-learning detection systems.