Peach fruit real-time recognition in complex orchard environment based on improved YOLOv4-tiny

In order to achieve real-time peach fruit recognition in complex orchard environments, a real-time peach fruit recognition approach（YOLOv4-tiny-Peach） was proposed based on YOLOv4-tiny. The method firstly optimized the feature information in its channel dimension and spatial dimension by introducing a Convolution Block Attention Module（CBAM） in the backbone network, then, it improved the accuracy of small target recognition by adding a large-scale feature layer（F3） to the neck network, and finally, it used a Bidirectional Feature Pyramid Network（BiFPN） to fuse feature information at different scales. Through training and comparison, the average accuracy AP of YOLOv4-tinyPeach model was 87. 88%, the accuracy P was 91. 81%, the recall R was 73. 84%, and the F1 score was 81. 85%under the test set. AP was increased by 5. 46%, P by 2. 29%, R by 4. 09%, and F1 by 3. 44%, respectively, compared to before the improvement. The recognition of peach fruit images under different numbers, different ripening stages, and occlusion was performed in order to verify the adaptability of the improved model in the complex environment of orchards, and the recognition effect of the improved model was compared with that of the original model. The results showed that the recognition accuracy of YOLOv4-tiny-Peach was higher than that of the YOLOv4-tiny in all three cases, especially in scenes with a wide field of view and unripe fruit. The YOLOv4-tiny-Peach model uses memory of 27. 4 MB and has a inference speed of 49. 76 fps, making it more suitable for embedded agricultural equipment. The method provides real-time and precise target identification assistance for autonomous peach fruit picking under complex environment in orchards.