Urban waste management is one of the major environmental and economic challenges in modern cities. A considerable part of municipal solid waste consists of recyclable materials such as cardboard, glass, metal, paper, and plastic. Accurate detection and separation of these materials can improve recycling quality and reduce the environmental burden of urban waste. This study proposes a deep learning-based computer vision framework for intelligent detection and separation of recyclable urban waste. Unlike single-label image classification methods, the proposed framework performs multi-instance object detection and generates a bounding box, class label, and confidence score for each detected item. The architecture includes preprocessing and augmentation, lightweight feature extraction, a combined attention module, multi-level feature fusion, and a final detection stage. The dataset contains 4250 images and 6815 annotated waste instances. The proposed method achieved Precision, Recall, F1-score, mAP@0.5, and mAP@0.5:0.95 values of 0.934, 0.919, 0.926, 0.947, and 0.742, respectively. The results indicate that the proposed framework can support automated waste sorting systems, smart bins, robotic separators, and recycling facilities.