When the picking robot picks green peaches, there are problems such as the color of the fruit being similar to the background color, overlapping fruits, and small fruit size, uneven lighting, and branches and leaves occlusion. As a result, the picking robot cannot quickly detect green peaches. In order to solve the above problems, a lightweight object detection network for fast detection of green peaches is proposed, which is composed of a backbone network, feature enhancement network, Lightweight Self-Attention (LSA) network, and four-scale prediction network. First, the lightweight detection unit LeanNet of the backbone network is designed, which uses the idea of deep separable convolution to achieve fast detection. Secondly, the feature enhancement module (P-Enhance) is designed, which uses convolution kernels of different receptive fields to extract different perceptual information in the feature map, which enhances the network’s feature extraction ability for green peach. Then, the LSA module is designed to generate a local saliency map based on green peach features, which effectively suppressed the irrelevant area of the branch and leaf background. Finally, a four-scale prediction network is designed, in which the Four-scale Pyramid Fusion (FSPF) module can generate a four-scale feature pyramid, which includes the color and shape of the green peach at different network depths, and is conducive to the detection of small volume green peaches. The experimental results show that precision, recall, and F1 of our method in the green peach test set reached 97.3%, 99.7%, and 98.5%, respectively. In the actual picking scenes, Qualcomm Snapdragon 865 embedded devices equipped with different state-of-the-art methods are used. Through comparative experiments in various scenarios, compared with the state-of-the-art method, both in terms of experimental data and visual effects, there is a significant improvement, which can meet the real-time object detection needs of picking robots.
