Multi-scale Convolution Kernels and Residual Components for Deciduous Leaves Detection

The paper investigates the efficiency of Yolo v5 in terms of four versions, including, Yolo v5s, Yolo v5m, Yolo v5l, Yolo v5x, for small object detection, by discussed the structures of two main components in Yolo v5, that is, Backbone for feature extraction and PAnet for feature fusion, thereafter, the various numbers of the convolution kernels and the residual components embedded into four versions of Yolo v5 structures are elucidated. The experiment exhibits a representative image contained the numerous deciduous leaves covered by the randomly rendering sunshine and the shadowing spots affected by the giant trees, thereafter, validated by four versions of Yolo v5. Finally, the training cost regarding the tradeoff between the epoch number and the training loss are reported that probably applied to the embedded system for small-object detection applications.