CSP-Flexi-Net: A Backbone Concept with Flexible Cross-Stage Partial Network Architecture for Object Detection

Reuben Zenas Jenn Wei Soh

Deep learning has been a subject of great interest in recent times. There has been much progress in the development of many deep learning-based applications. The recent trend is to implement these deep learning-based applications onto edge devices. One such application is object detection, which is a fundamental challenge in computer vision. To successfully implement object detection frameworks on edge devices, it is necessary to design smaller and more efficient backbone frameworks. A key component of object detection models is the backbone, which is a convolutional neural network that performs most of the computations of the object detection process. To reduce the computational costs in the backbone, the convolutional neural network needs to be compact and efficient. However, the trade-off in designing smaller and more efficient convolutional neural networks is the degradation in model performance. The challenge is to design a compact and efficient convolutional neural network while maintaining an acceptable level of model performance compared to a larger convolutional neural network. In this thesis, new variations of the convolutional neural network, CSPResNet (Cross Stage Partial Residual Network), are proposed and evaluated. These are created based on two backbone structures: the CSPResNet models with Basic Blocks and the CSPResNet models with Bottlenecks. Experiments were conducted for both backbone structures to compare their object detection performances. It was found that the CSPResNet18C model designed with Basic Blocks had better accuracy by 2.9%, reduced parameters by 61.5%, and reduced model size by 47.3% as compared to the CSPResNet50 model designed with Bottlenecks. In addition, we also introduced a backbone concept, the CSP-Flexi-Net, and its associated design strategy based on a parent CSPResNet architecture. This is a composite backbone structure concept where different backbone blocks can be substituted within a parent CSPResNet to reduce its computational costs and model size. Two variants were explored: i) ResNet Basic Blocks and Bottlenecks mixed and ii) ResNet and Depthwise Separable Convolutional (DSC) Basic Blocks mixed. Both architectures show reduced computational costs with mixed results in performance. For the Composite CSPResNet50, it shows 5.4% improvement on its mAP-50 as compared to the current existing original CSPResNet50. For the DSC-composite CSPResNet18C, it shows 3.2% drop on its mAP-50 but with 21.7% lesser parameters. These results demonstrate a slight trade-off in performance but with reduced computational cost. Overall, these experimental results demonstrate the potential of CSP-Flexi-Net as a viable alternative technique for model compression as compared to other existing techniques, such as neural architecture search (NAS), knowledge distillation, quantisation, and pruning.

CSP-Flexi-Net: A Backbone Concept with Flexible Cross-Stage Partial Network Architecture for Object Detection

(1)