To overcome the limitations of commercial embedded computing resources and achieve real-time performance, high energy efficiency, and deployment capability in complex industrial environments for object detection tasks, this study proposes a lightweight object detection algorithm optimization scheme. First, a multi-branch depthwise separable convolutional module is developed to replace the conventional convolution in the backbone network, and multi-scale feature representation is achieved by introducing a multi-scale pyramid-shaped module. Second, an adaptive spatial feature fusion strategy is adopted to maximize the multi-scale fusion of detection head features, and channel compression significantly reduces the computational load. Finally, global sensitivity analysis is used for layer pruning and channel pruning to further simplify the model. The proposed algorithm achieves a running speed of 41.3 FPS on embedded systems and, under the theoretical conditions of reducing the number of parameters to 2.2 million and the computational cost to 2.28 GFLOPs, achieves an average accuracy of 72.1%. This demonstrates that a good balance between detection effectiveness and accuracy can be achieved under limited resources.
Xu et al. (Thu,) studied this question.