Tomato fruit detection and phenotype calculation method based on the improved RTDETR model

Rapid detection of tomato fruits and accurate acquisition of phenotypic traits are of great significance for robotic automatic picking control, yield prediction, and variety breeding. Tomato fruits are often densely distributed in a complex canopy and obscured by branches and leaves, making it difficult to accurately detect tomato fruits and obtain phenotypic traits without damage. This paper proposes an automatic detection method for tomatoes based on an improved RTDETR model. Firstly, on the basis of the self-made calibration plate, the color image sensor is used to acquire the tomato image. Then, a CASA structure consisting of three modules: Multiscale Dilated Convolution (MDC), Focused Feature Downsampler (FFD) and Adaptive Feature Upsampler (AFU) was designed and embedded into the Neck structure of the RTDETR network to construct a tomato fruit detection method based on the improved RTDETR model. Finally, by integrating machine learning and graphics processing technology, a fruit color extraction method was established based on the CIELAB color space, a fruit diameter calculation method based on edge detection and Hough transform, and a fruit weight and circumference measurement method based on statistical regression models. The experimental results show that the $mAP\_0.5$ of the tomato fruit detection model established in this paper reaches 0.86, which is 3% higher than the original model; The correlation coefficient between the calculated and measured values of the horizontal and vertical diameters of the fruit was 0.79, and the mean square error ($\mathrm{MSE}$) of the weight and circumference of the fruit was 0.26 and 0.27, respectively. This achievement has realized an accurate, lossless, and fast method for tomato fruit detection and phenotype calculation, providing quantitative reference indicators for fruit detection, positioning, and control of tomato automatic picking robots, and can provide technical support and guarantee for crop yield prediction and variety breeding.