MCBPnet: An efficient lightweight apricot recognition model

With the continuous progress of agricultural modernization, the application of automation and intelligent technologies in agricultural production has become increasingly widespread. Green apricots, as a significant economic crop, present unique challenges and opportunities for precision management. Innovation in fruit recognition and detection technologies is critical for enhancing the accuracy and efficiency of crop management. This study developed a novel target recognition model, the MCBPnet model, leveraging advanced deep learning techniques to address the limitations of traditional methods under complex environmental conditions. By doing so, it aimed to provide an efficient and reliable solution for the automatic detection of green apricot fruits.The core of this research lay in integrating the CBAM (Convolutional Block Attention Module) into the MobileNetV3 architecture. This addition significantly enhanced the model's ability to focus on the most information-rich areas within an image, thereby improving its precision in identifying apricot fruits. By prioritizing regions of an image that are most relevant to fruit detection, the model effectively allocated more computational resources to features closely associated with the task. As a result, a new model structure, IRCBAM (Inverted Residual Convolutional Block Attention Module), was developed and incorporated as the backbone network of MCBPnet. To further enhance the model's performance, the BiFPN (Bi-directional Feature Pyramid Network) was introduced into the neck network. By constructing a feature pyramid, BiFPN enabled the model to capture multi-scale features from images, which was essential for detecting fruits of varying sizes within a single frame. This multi-scale feature representation allowed the model to maintain a high detection rate across a wide range of fruit sizes and orientations. Moreover, the study addressed challenges posed by occlusions, such as fruits being partially obscured by leaves or other fruits. For this purpose, the PConv(Partial Convolution) structure was utilized in the detection head. This method ensured that the model maintained detection accuracy even when only a portion of a fruit was visible. By leveraging PConv, the model achieved reliable recognition under conditions of partial visibility, thereby increasing its accuracy and robustness. The results of this study demonstrated that the MCBPnet model achieved significant advantages across key performance metrics, including mAP50 and mAP50-95. Specifically, the model attained a precision of 0.988, a recall of 0.985, an mAP50 of 0.994, and an mAP50-95 of 0.968, reflecting its high detection accuracy. Furthermore, compared to the YOLOv8 model, the MCBPnet model reduced parameter size and FLOPs by 2.702% and 24.691%, respectively. This lightweight design ensured that the model could be deployed on edge devices, such as drones or handheld scanners, without requiring energy-intensive computational resources. In summary, the high-performance metrics achieved by the MCBPnet model validated the effectiveness of the integrated techniques. Its ability to maintain high accuracy across various detection indicators represents a significant advancement in agricultural automation. The development of the MCBPnet model marked an important step toward automating apricot fruit detection. Its combination of advanced deep learning techniques and lightweight design makes it a powerful tool for improving the efficiency and accuracy of apricot harvesting. By providing technological support for the automated recognition and picking of green apricots, this model contributes to the broader adoption of modern, intelligent, and sustainable crop management practices. As agriculture continues to embrace technological advancements, models like MCBPnet will play a crucial role in driving the transition toward more efficient and intelligent agricultural practices.