Network Model for Multi-Scale Dense Object Detection Focused on Small Objects

Deep learning methods have been proven to surpass previously state-of-the-art machine learning techniques in several fields. As a fundamental problem in computer vision, object detection has gained widespread attention in recent years. However, detecting small objects, multi-scale targets, and densely packed items remains challenging due to the following inherent reasons: (1) Model weight updates are often affected by noisy data, variations in training samples, and gradient oscillations; (2) Small objects occupy fewer pixels, and after multiple down-sampling and convolution operations, their features become increasingly blurred or lost; (3) Traditional IoU loss functions only consider the overlapping area between the predicted and ground truth boxes, ignoring factors such as the shape, ratio, and orientation of the boxes. To address the first issue, we propose an improved attention mechanism module (CEMA) and introduce a multi-scale convolutional neural architecture (InceptionNext). To tackle the second issue, we add a parallel small object detection layer. Compared to baseline methods, our improved model shows significant enhancements in object detection, especially in complex multi-scale environments with small targets captured by drones.