Abstract
Abstract
Accurate and automated coffee bean classification is important for improving consistency and efficiency in agricultural quality assessment. When applied to fine-grained agricultural classification tasks with constrained dataset footprints, standalone deep convolutional neural network (CNN) architectures frequently experience performance degradation and severe overfitting because unguided neural layers find it difficult to self-optimize macro-geometric topologies. This research offers a unique asymmetric Hybrid Feature Fusion Framework to resolve this engineering bottleneck by integrating high-purity handcrafted morphological metrics with abstract deep texture representations. The proposed architecture functions in three separate phases: first, a geometric modeling baseline generates a 7x1 morphological feature vector, resulting in a standalone Random Forest accuracy of 93.95%. Subsequently, a standalone CNN captures localized spatial variations, attaining an accuracy of 83.59%. Third, these feature spaces are connected by an asymmetric Horizontal Concatenation Fusion Pipeline, which creates a combined 39x1 super vector by embedding the geometric parameters next to 32 abstract deep texture vectors from the CNN's penultimate dense layer. The resultant Ensemble Hybrid Classifier effectively reduced representation noise, improving standalone neural performance by 2.41% and achieving a stabilized system accuracy of 86.00% with consistent predictive stability across genetic profiles when tested on a validation cohort of 463 regional coffee bean samples. The whole hybrid inference matrix was then implemented as a lightweight edge prototype using an embedded ATmega2560 microcontroller architecture with hardware in the loop liquid crystal display verification in order to confirm industrial feasibility. According to experimental findings, the hybrid framework effectively introduces structural domain knowledge into deep feature spaces, providing a highly generalizable and computationally effective route for inexpensive, decentralized edge-sorting systems.