Diabetic retinopathy (DR) is a leading cause of blindness globally, requiring timely detection and classification to prevent vision loss. Deep learning techniques offer significant potential for automating DR detection by analyzing retinal fundus images with high precision. This paper proposes a RetinoNet model that consists of MobileNetV3, Convolutional Block Attention Module (CBAM), Atrous Spatial Pyramid Pooling (ASPP), and Feature Pyramid Network (FPN). MobileNetV3 provides a lightweight and efficient foundation for feature extraction, while CBAM emphasizes critical spatial and channel information, enabling the detection of subtle retinal abnormalities. ASPP captures multi-scale contextual information through atrous convolutions, improving the model's ability to identify lesions of varying sizes and shapes. FPN combines hierarchical features from multiple network levels, ensuring both fine-grained details and high-level semantics are leveraged for accurate classification. The model was trained on the APTOS dataset. Evaluation metrics such as accuracy, precision, recall, and F1 score demonstrate the effectiveness of the proposed model in achieving state-of-the-art performance for DR detection and classification across five severity levels. This approach addresses computational challenges and improves generalization, making it suitable for both clinical and remote healthcare applications.