PC-CSP: An Efficient Partial Convolution in Cross Stage Partial Network for Small Targets Detection

The common approach for small object detection does not work well in complex scenes with objects of multiple sizes. This is because the detection of small objects and objects of other sizes often difficult to take into account at the same time. To overcome these limitations, we introduce an efficient Partial Convolution in Cross Stage Partial (PC-CSP) network for detecting small targets. First, the introduction of the DCNv3 module improves the feature extraction for detection, slimneck is utilized to enhance the attention mechanism suitable for objects of varying sizes. Then, PC-CSP is employed to streamline the detection backbone network. Last, the loss function is refined by incorporating the concept of Inner. Through extensive experiments, our method demonstrates a significant increase in accuracy of 6.8%, and [email protected] by 8.97% on the standard dataset compared to the original model. Overall, the proposed method exhibits excellent performance for small objects, such as: mobile phones, food, and faces.