关键词: 油茶果, YOLOv4-Tiny网络, 深度学习, 图像识别

Abstract: Aiming at the problem of wrong detection and missing detection of Camellia oleifera fruit in natural environment due to the factors such as occlusion, overlap of branches and leaves, similar color of fruit and leaf, a recognition method of Camellia oleifera fruit in natural environment based on improved YOLOv4-Tiny was proposed. A largescale output feature layer and a pyramid pooling model were introduced into the backbone network, through which the network performance caused by the insufficient representation ability of the model was solved when the network was deepened due to occlusion or overlapping Camellia oleifera fruit. The Kmeans algorithm was used to cluster a priori box suitable for the selected data set to improve the model detection effect. The experimental results showed that the recall rate of the improved YOLOv4-Tiny algorithm on 1 901 test sets was 86.15%, the mAP value was 94.19%, and the F1 value was 0.9, which were 1.2, 0.34 and 2 percentage points higher than before. After the improvement, the average detection time of the algorithm in a single image was 0.025 s, and the detection speed was 40.45 f/s, which was 7.41% less than that before the improvement, and the detection speed was increased by 3.87%. The recognition accuracy of the algorithm in this paper was 4.36 and 1.55 percentage points higher than that of the autoencoder and the convex hull theory, respectively. The single image detection time of FasterRCNN, autoencoder and convex hull theory was 8.4, 66.4 and 19.64 times of the algorithm in this paper, respectively. It was proved that the algorithm had a good effect on the recognition of Camellia oleifera fruit images in complex natural environment, and met the requirements of realtime picking.

Key words: Camellia oleifera fruit, YOLOv4-Tiny network, deep learning, image identification