Evolutionary channel pruning for real-time object detection

Real-time object detection plays a crucial role in edge devices applications. Pruning methods are usually used to effectively eliminate redundant parameters of the object detection network so that it can detect objects efficiently. However, traditional pruning methods often result in a significant drop in accuracy, requiring time-consuming fine-tuning to restore the accuracy of the network. To address this issue, we propose an evolutionary channel pruning (ECP) method to reduce the redundant parameters in the network. Our proposed ECP method effectively reduces parameter redundancy and computation complexity in object detection networks while maintaining detection accuracy. Additionally, we introduce a novel Channel Information Mixing Convolution (CIMConv) that leverages more cost-effective operations to achieve higher accuracy and reduce the complexity associated with standard convolution. By applying our proposed ECP and CIMConv to the existing object detection methods, we achieve a superior balance between accuracy and complexity compared to state-of-the-art detectors. Notably, on challenging public datasets such as GTSDB, S2TLD, TT100K, Wider Face, and Microsoft COCO, our proposed ECP substantially decrease the number of parameters and FLOPs of YOLOv5, simultaneously improving detection accuracy.