Apple_Yolo: Apple Detection Method Based on Channel Pruning and Mixed Distillation in Complicated Environments

Abstract

Rapid and precise positioning of apples, along with intelligent detection, play a pivotal role in the process of picking apples. Nevertheless, the existing crop detection methods that rely on deep learning sometimes require substantial computational resources and memory, consequently limiting their feasibility for mobile device implementation. This study presents a lightweight algorithm technique for detecting apple targets to address the problem of insufficient storage space and restricted computational capacity in apple-picking mobile devices. The method offers two distinct schemes based on different computing resources. The procedure consists of two primary phases. In the first stage of the lightweight process, the lightweight Feature Pyramid Network (LFPN) replaces the original trunk, followed by the utilization of lightweight down-sampling convolution (LDConv) to substitute the redundant convolutions in the trunk to reduce the number of parameters. Then, the Lightweight multi-channel attention mechanism (LMCA) is embed between the backbone network and the neck network to minimize the effects of unnecessary background. Finally, the model is distilled for the first time using mixed distillation to enhance the model's detection performance further. In the second stage of the lightweight, the Group_slim channel pruning is used to reduce redundant channels further. Subsequently, hybrid distillation is employed again to restore the accuracy of the pruning model. The results show that the average precision (AP) of the model presented in this study is 1% higher than that of the baseline model, given that the parameter count is only about 800k. The models of both schemes can achieve an inference speed of over 17 frames per second on the central processing unit(CPU).

Keywords: LMCA, LFPN, LDConv, Group_slim, Distillation