In order to improve the universality and practicality of plant chlorophyll detection equipment, this study investigates whether there is a fitting relationship between RGB images captured by mobile phones and microcontrollers and the chlorophyll content of plant leaves. The purpose is to determine the relevant experiments that can predict chlorophyll through image processing, providing experimental basis for future dynamic nondestructive detection of plant chlorophyll based on deep learning. By using OpenCV to extract the region of interest (RoI) from the image and applying mean filtering, Gaussian filtering, and median filtering, the original image and the three filtered images are subjected to threechannel color feature separation. The least squares method (LS) is used to perform fitting analysis on various combinations of color feature parameters and measured chlorophyll values. It is found that the fitting effect of mean filtering is generally better among the four types of images. In mean filtering, the image captured by the mobile phone K40 has a feature combination of (B-G-R)/(B+G) with a coefficient of determination of 0.912 for fitting the leaf chlorophyll. The image captured by the microcontroller ESP32_CAM has a feature combination of (G-B)B/(R+G) with a coefficient of determination of 0.778 for fitting the leaf chlorophyll. By applying gradient operation to iteratively process the RoI of mean filtering, it is found that the coefficient of determination of K40 slightly decreases, while the coefficient of determination of ESP32_CAM improves. Through the verification of prediction models for K40 and ESP32_CAM, both show that the random forest (RF) regression model performs the best. In K40, the coefficient of determination for the training set is 0.953, the root mean square error for the training set is 1.161, the coefficient of determination for the prediction set is 0.930, and the root mean square error for the prediction set is 1.516. In ESP32_CAM, the coefficient of determination for the training set is 0.794, the root mean square error for the training set is 2.510, the coefficient of determination for the prediction set is 0.695, and the root mean square error for the prediction set is 2.985.