Research on tomato detection in natural environment based on RC-YOLOv4

Abstract

In natural environment, the factors such as illumination change, background interference and leaf occlusion have a great impact on the tomato detection accuracy of the picking robot. To address this problem, based on YOLOv4, a new backbone network R-CSPDarknet53 is constructed by fusing the residual neural network to establish the jump connection between the front and back layers to ensure the incomplete loss of low-dimensional small target features. In addition, by replacing the maximum pool in the original SPP network with the deep separable convolution model, C-SPP is proposed to realize feature information reuse and multi-scale fusion. On this basis, a tomato detection model RC-YOLOv4 is constructed, which improves the detection accuracy of tomato in natural environment. The test results show that the tomato detection accuracy and recall rate of RC-YOLOv4 model in natural environment are 88% and 89% respectively, the average detection accuracy is 94.44%, the harmonic mean F1 is 0.89, and the video stream detection speed is 10.71 frames/s. Compared with the original model, RC-YOLOv4 improves the accuracy by 1.52% with little loss of speed, and has stronger environmental adaptability. It has strong robustness under different occlusion degrees and illumination conditions, and can better adapt to tomato detection in complex natural environment.