How Does Spectral Technology Make Fruit Quality Sorting Smarter?

水果的品質(zhì)分選直接決定了供應(yīng)鏈的經(jīng)濟(jì)價值，精準(zhǔn)識別內(nèi)部損傷不僅能顯著降低優(yōu)質(zhì)果的損耗率，更能提供穩(wěn)定可靠的高品質(zhì)水果。然而，傳統(tǒng)的人工分選和外部檢測難以發(fā)現(xiàn)瘀傷、凍傷、水心病等內(nèi)部缺陷，導(dǎo)致大量外表完好的水果因隱性損傷而被誤判。

光譜與高光譜成像技術(shù)的出現(xiàn)，讓水果內(nèi)部品質(zhì)的無損檢測成為可能。這些技術(shù)通過解析水果內(nèi)部水分、糖分及細(xì)胞結(jié)構(gòu)的特征光信號，實現(xiàn)對瘀傷、凍傷、海綿組織病變等隱性缺陷的“透視"，從而大幅提升分選精度。

Fruit quality sorting directly determines the economic value of the supply chain. Accurately identifying internal damage not only significantly reduces the loss rate of premium fruits but also provides stable and reliable high-quality fruits for the premium market. However, traditional manual sorting and external inspection struggle to detect internal defects such as bruises, frost damage, and watercore, leading to misjudgment of large quantities of outwardly intact fruits due to hidden damage.

The emergence of spectral and hyperspectral imaging technologies has made non-destructive testing of internal fruit quality possible. These technologies analyze characteristic optical signals related to internal moisture, sugar content, and cellular structure, enabling "visualization" of hidden defects like bruises, frost damage, and spongy tissue disorders, thereby significantly improving sorting accuracy.

水果分選線，圖片源自網(wǎng)絡(luò) / Fruit sorting line, image source: Internet

在國內(nèi)外，科研團(tuán)隊已通過大量實驗驗證了光譜技術(shù)在水果分選中的潛力。例如，江蘇大學(xué)團(tuán)隊利用高光譜成像系統(tǒng)檢測蘋果的輕微損傷，發(fā)現(xiàn)547nm波段的特征光譜能清晰反映皮下細(xì)胞損傷，通過主成分分析（PCA）提取該波段圖像，結(jié)合二次差分算法消除果面亮度不均干擾，最終實現(xiàn)88.57%的損傷識別率。

Globally, research teams have validated the potential of spectral technology in fruit sorting through extensive experiments. For example, a team from Jiangsu University used a hyperspectral imaging system to detect slight damage in apples. They found that the characteristic spectrum at the 547nm wavelength clearly reflects subcutaneous cellular damage. By extracting images of this wavelength through principal component analysis (PCA) and combining it with a second-order difference algorithm to eliminate interference from uneven surface brightness, they achieved an 88.57% damage recognition rate.

蘋果的輕微損傷和正常區(qū)域的光譜曲線

Refectance spectra from the subtle bruise and normal region on the apple

類似地，國外研究團(tuán)隊在芒果海綿組織檢測中，通過優(yōu)化特定波段的Fisher特征選擇算法，使分類準(zhǔn)確率達(dá)到84.5%，且預(yù)測缺陷位置與實際損傷的誤差小于1mm。

Similarly, a foreign research team optimized the Fisher feature selection algorithm for specific wavelengths in mango spongy tissue detection, achieving a classification accuracy of 84.5% with prediction errors of defect locations within 1mm of actual damage.

缺陷樣本與健康樣本的光譜圖，(a) 波長范圍673nm–1100nm，(b) 波長范圍1100nm–1900nm

A plot of defective and healthy samples  (a) Wavelength range 673 nm–1100 nm. (b) Wavelength range 1100 nm–1900 nm.

這些研究成果為實際產(chǎn)線應(yīng)用奠定了基礎(chǔ)?；诠庾V的水果分析系統(tǒng)通常由光源模塊、光譜儀/高光譜成像儀、傳送帶等核心硬件組成，其工作流程包括樣品采集、光譜數(shù)據(jù)預(yù)處理、化學(xué)分析方法測定水果樣品成分的準(zhǔn)確含量、模型構(gòu)建與驗證、優(yōu)化模型等關(guān)鍵步驟。而在實際分選場景中，這個流程如何高效運(yùn)行？關(guān)鍵在于自動化和光譜檢測的緊密結(jié)合：

1. 動態(tài)觸發(fā)：傳送帶水果抵達(dá)檢測位，光電傳感器觸發(fā)光源

2. 光譜采集：光源發(fā)射出光，光譜儀獲反射/透射光譜

3. 數(shù)據(jù)處理：光譜儀分解特征峰，分析模型實時輸出糖度/損傷值

4. 分揀執(zhí)行：觸發(fā)品質(zhì)分級

These research findings lay the foundation for practical production line applications. Spectral-based fruit analysis systems typically consist of core hardware such as a light source module, spectrometer/hyperspectral imager, and conveyor belt. The workflow includes key steps such as sample collection, spectral data preprocessing, chemical analysis to determine the accurate content of fruit sample components, model construction and validation, and model optimization. In actual sorting scenarios, the efficiency of this process hinges on the seamless integration of automation and spectral detection:

1. Dynamic Triggering: Photoelectric sensors activate the light source when fruit reaches the detection position on the conveyor belt.

2. Spectral Acquisition: The light source emits light, and the spectrometer captures the reflected/transmitted spectra.

3. Data Processing: The spectrometer decomposes characteristic peaks, and the analysis model outputs real-time brix/damage values.

4. Sorting Execution: Quality grading is triggered for sorting.

高光譜系統(tǒng)的示意圖（由于水果的尺寸大小、果肉薄厚，糖酸度高有低，且分布不均的情況，光譜采集時光源擺放有多種方式）

Schematic diagram of a hyperspectral system (Due to variations in fruit size, flesh thickness, and uneven distribution of sugar/acid content, multiple light source configurations are used during spectral acquisition)

目前，光纖光譜儀因其成本低、結(jié)構(gòu)緊湊等優(yōu)勢，仍是水果分選的主流設(shè)備。但對于圣女果、櫻桃等小尺寸水果，光纖光譜儀的檢測效率可能受限，而高光譜成像儀憑借其空間與光譜信息的同步獲取能力，理論上能實現(xiàn)更高效的分選。然而，高光譜設(shè)備的成本和數(shù)據(jù)處理復(fù)雜度仍是實際應(yīng)用中的挑戰(zhàn)。

未來，隨著硬件優(yōu)化和算法的持續(xù)升級，光譜技術(shù)有望在更多水果品類中實現(xiàn)高效、經(jīng)濟(jì)的分選方案，推動水果供應(yīng)鏈向更智能、更精準(zhǔn)的方向發(fā)展。

Currently, fiber optic spectrometers remain the mainstream equipment for fruit sorting due to their low cost and compact structure. However, for small-sized fruits like cherry tomatoes and cherries, the detection efficiency of fiber optic spectrometers may be limited. In contrast, hyperspectral imagers, with their ability to simultaneously capture spatial and spectral information, theoretically enable more efficient sorting. Nevertheless, the cost of hyperspectral equipment and the complexity of data processing remain challenges in practical applications.

Looking ahead, with continuous hardware optimization and algorithm advancements, spectral technology is expected to deliver efficient and cost-effective sorting solutions for more fruit varieties, driving the fruit supply chain toward smarter and more precise development.

案例來源 / Source：

1. Zhao, J.-W., Liu, J.-H., Chen, Q.-S., & Vittayapadung, S. (2008). 利用高光譜圖像技術(shù)檢測水果輕微損傷 [Detection of slight fruit bruises using hyperspectral imaging technology]. Transactions of the Chinese Society for Agricultural Machinery, 39(1), 106-109.

2. Raghavendra, A., Guru, D. S., & Rao, M. K. (2021). Mango internal defect detection based on optimal wavelength selection method using NIR spectroscopy. Artificial Intelligence in Agriculture, 5, 43-51.