Measurement of Seed Cotton Color Using RGB Imaging and Color-Unet

Abstract

Color is a key indicator in evaluating seed cotton quality. Accurate and rapid detection of seed cotton color is essential for its storage, processing, and trade. In this study, an RGB imaging and semantic segmentation-based method was proposed for seed cotton color detection. First, a color detection system utilizing machine vision technology was developed to capture seed cotton images. Next, a Color-Unet model, incorporating convolutional block attention and improved inception E modules based on Unet, was applied to effectively remove impurities and shadows from the images, resolving the over-segmentation issue commonly encountered in traditional threshold segmentation. The results demonstrated that the pixel accuracy of segmentation reached 97.20%, the mean intersection over union was 91.81%, and the average segmentation speed was 322.3 ms per image. The Color-Unet model effectively addressed the over-segmentation problem. Subsequently, seed cotton color indexes were calculated using Hunter color formulas based on the segmented images. To evaluate the accuracy of color measurement obtained with the proposed method, a regression analysis was performed, comparing the results of those from the HX-410 measurement. The coefficient of determination of yellowness was 0.883, with a root mean square error of 0.150 and a mean relative error of 2.61%. The coefficient of determination of reflectance degree was 0.832, with a root mean square error of 1.56% and a mean relative error of 1.84%. The proposed method allows for the rapid and accurate assessment of seed cotton color from RGB images, providing a valuable technical reference for seed cotton color evaluation.

1. Introduction

Seed cotton is a key agricultural raw material in the textile industry and plays a significant role in global agricultural trade [1]. Its quality directly influences its economic value, making quality assessment a crucial step in international seed cotton trade [2]. Among various quality indicators, color is paramount for evaluating seed cotton quality and determining market price [3]. Color assessment forms the basis for grading seed cotton, reflecting its storage and processing conditions [4,5]. Therefore, the accurate and rapid detection of seed cotton color is essential for fair trade and optimal processing utilization. Typically, cotton color indices are expressed in terms of yellowness (+b) and reflectance degree (Rd) within Hunter Lab color coordinate system [6]. Nickerson et al. [7] developed an automatic cotton colorimeter that employed photoelectric conversion to detect lint cotton color, followed by the High-Volume Instrument (HVI), which is still in use today. However, these instruments cannot be directly applied to seed cotton color detection, as they are specifically designed for evaluating lint cotton color [8]. Lint cotton, having fewer impurities, allows these instruments to operate effectively based on photoelectricity principle. They measure the diffuse light reflected from the sample surface and subsequently calculate the color value. In contrast, seed cotton often contains more impurities [9]. The presence of impurities such as boll shells, stems, and seeds on the sample surface can create visible shadows, complicating the color measurement process. Consequently, the calculated color value reflects the mixture of impurities, shadows and cotton fiber [3,10]. Therefore, shadows and impurities present significant challenges in accurately measuring the seed cotton color [11,12]. Currently, human visual assessment is the primary method for evaluating seed cotton color. However, this approach is time-consuming, labor-intensive, and inefficient. Thus, there is an urgent need for a rapid, accurate method for evaluating seed cotton color that effectively mitigates the effects of shadows and impurities.

Machine vision technology (MVT) offers significant advantages in object recognition, characterized by stability, speed, and high precision. In recent years, MVT was applied to cotton color detection. For instance, Cui et al. [13] utilized MVT to measure variations in lint color, achieving a correlation coefficient of 0.99 between MVT results and those obtained from the HVI, indicating MVT’s effectiveness in detecting surface color changes. However, this study did not account for the effects of shadows and impurities. Guo et al. [14] developed an MVT-based detection device for lint color and investigated the influence of varying pressure on measurement accuracy, finding that a pressure of 40 N minimized error rate, with Rd error rates ranging from 0.06% to 0.60% and +b error rates ranging from 0.35% to 4.09%. Nevertheless, the effects of impurities and lint density were not considered, suggesting potential for further error reduction by mitigating impurity effects. Heng et al. [15] enhanced the correlation of lint color measurements (correlation coefficient > 0.92) by using threshold segmentation to remove impurities from lint images, although shadows were not addressed. Li, H. et al. [10] proposed a Secondary Dynamic Threshold Segmentation algorithm based on Multichannel Fusion (SDTS-MF) algorithm for segmenting impurity and shadow pixels in seed cotton images, significantly improving the correlation of color detection indexes for seed cotton, with R² values for +b and Rd increasing by 0.243 and 0.448, respectively. However, this algorithm, built on conventional threshold segmentation, lacks generalizability, especially with yellow-stained cotton fibers, where it often results in over-segmentation.

Semantic segmentation networks (SSNs) based on deep learning are widely employed in image segmentation due to their strong generalization and feature extraction capabilities. SSNs can extract complex image features that are challenging to manually define, enabling pixel-level image classification. These networks found extensive applications in agriculture [16,17,18]. For instance, Wei et al. [19] applied a modified SSN to segment foreign fibers within cotton images, achieving an accuracy of 96%. Similarly, Xu et al. [20] developed the GlandSegNet model to segment pigment glands in cotton leaf images, reaching an accuracy of 0.951. These results demonstrated the effectiveness of SSNs for cotton image segmentation. Unet, a classical model for semantic segmentation, showed excellent results in agricultural image segmentation [21,22]. However, Unet’s use of numerous convolutions and down-sampling and up-sampling operations results in a large number of training parameters, increasing computational demands, and makes training and segmentation time-consuming. Therefore, optimizing the model is essential for achieving fast and accurate seed cotton color evaluation.

In summary, this study proposed a novel approach for seed cotton color detection using RGB imaging and SSNs. The objectives of this research were as follows: (1) to develop a detection system capable of rapidly and accurately capturing seed cotton images and measuring color indices; (2) to construct a Color-Unet model to enhance the accuracy and speed of shadows and impurities segmentation in seed cotton images, addressing over-segmentation issues; and (3) to evaluate the effectiveness of Color-Unet model. The specific research process is illustrated in Figure 1. This study provided a valuable technical reference for enhancing seed cotton color detection.

2. Materials and Methods

2.1. Samples

Seed cotton samples were collected from Shihezi General Farm of Shihezi city in 2022. Figure 2 shows the location of Shihezi General Farm. A total of 45 bales of seed cotton (1 kg each), varying in color grades and varieties, were gathered during the seed cotton harvest (Figure 3). First, sand and dust were removed using a separator to prevent color contamination. Then, 350 g of seed cotton sample was weighed from each bale and labeled accordingly. These cotton samples were used for color measurement, while the remaining seed cotton sample (approximately 650 g) from each bale was reserved for image acquisition. All samples were subjected to standardized treatment in accordance with ASTM D5867.

2.2. Seed Cotton Color Detection System

A seed cotton color detection system based on MVT was developed. The system comprised a computer, a programmable logic controller (XC1-10R-E, Wuxi Xinjie Electric Co., Ltd., Wuxi, China), an image acquisition unit, and an image processing program (Figure 4). The image acquisition unit included an electric actuator (Maximum Thrust: 500 N, Wenzhou Pfeid Mechanical & Electrical Co., Ltd., Wenzhou, China), a digital camera (MV-CE060-10UC, Hangzhou Hikvision Digital Te. Co., Ltd., Hangzhou, China), a rectangular dark box (65 cm × 58 cm × 140 cm), four LED lamps (CRI: ≥ 90; CT: 6500 K; Shenzhen 3nh Lighting Co., Shenzhen, China), and a round sample box (d = 40 cm, h = 16 cm). The electric actuator, controlled by the programmable logic controller, provided consistent thrusts to compress the cotton using a transparent optical quartz glass plate. The four LED lamps, designed as standard surface light sources, illuminated the cotton sample at a 45° angle in accordance with ASTM D1729. The dark box, lined with black opaque sandpaper on its top and sides, minimized external light interference during image acquisition. The image processing and color index calculation program was developed using Python 3.9.

2.3. Dataset Preparation and Data Augmentation

A total of 450 images were obtained using the seed cotton color detection system from 45 packages of seed cotton samples. Before each image acquisition, the seed cotton samples were thoroughly mixed manually. After each acquisition, the samples were further mixed, with each sample captured 10 times to ensure uniformity. Prior to model training, the raw images were manually annotated to create a seed cotton dataset using LabelMe software (Version 4.5.13). In this process, cotton pixels were labeled with 0 (backgrounds), while impurity and shadow pixels were labeled with 1 (foregrounds), as shown in Figure 5. The dataset was then randomly divided into three sets: a training set (250 images), a validation set (130 images), and a test set (70 images).

Research demonstrated that deep learning models perform optimally with large training sample sizes [23]. To enhance the training set data, several data augmentation techniques were employed. During each training epoch, random cutting, horizontal/vertical flipping, and rotation (20°) were applied to the data. In each training epoch, 250 new training samples were generated, and training was conducted over 100 epochs. As a result, a total of 25,000 augmented images were used in the training process.

2.4. Image Segmentation Network

2.4.1. Unet Model

The Unet model is a widely used deep learning model for semantic segmentation [24], featuring a typical encoder–decoder structure network (Figure 6a). Unet can effectively extract highly related features through its convolutional layers and is known to perform well with small sample sizes [25]. However, the model’s numerous convolutional, down-sampling and up-sampling operations introduce a large number of parameters which significantly increase computational demands and extend training time [26]. To address this, a modified Unet model was proposed in this study for segmenting seed cotton images.

2.4.2. Color-Unet Model

To enhance the learning and generalization capabilities of the Unet model while reducing network parameters and run time, the number of convolutions was decreased, and the convolutional block attention module (CBAM) was integrated into the encoder. CBAM is a simple yet efficient attention module for convolutional neural network, offering strong capabilities in extracting both global and local features [27]. It sequentially infers attention features across spatial and channel dimensions, enabling adaptive feature refinement for feature maps (Figure 7a).

Depth and width significantly influence the performance of convolutional neural networks [28]. The Inception E module is particularly effective in convolutional decomposition by breaking down a large convolution into several smaller ones. To further leverage this concept, the 3 × 3 convolutional block in the Inception E module was decomposed into two smaller blocks, 1 × 3 and 3 × 1. Additionally, to better accommodate the varying shapes and scales of impurities and shadows in the seed cotton image, two asymmetric convolutions, 1 × 5 and 5 × 1, were introduced to replace the original 1 × 3 and 3 × 1 convolutions. The improved Inception E Module (IIEM) is depicted in Figure 7b. This approach drastically reduces the number of network parameters while expanding the network’s width and depth [29]. As a result, some convolutions in the decoder were replaced with the IIEM, forming the Color-Unet model, as illustrated in Figure 6b.

2.5. Model Training

The hardware used for model training included an Intel^® Xeon^® W-2133 CPU with 3.9 GHz frequency, an NVIDIA Quadro RTX 4000 GPU with 8 GB of VRAM, 128 GB of RAM, and 4 TB of ROM. The system ran on Windows 10. PyTorch version 2.0.1 utilized for model implementation, alongside cuDNN version 8.8 and CUDA version 11.8. Training was conducted using Python version 3.9.

Impurities and shadows were categorized together to minimize labeling complexity, simplifying the seed cotton image segmentation to a binary segmentation problem. Cross Entropy, a commonly used loss function in binary classification networks that measures the divergence between two probability distributions of the same random variable, was employed in this study [27]. The Cross Entropy was calculated according to Equation (1). The training process was iterative, utilizing the gradient descent method to accelerate convergence [25]. The Adam optimizer was used for network optimization, starting with a learning rate of 0.001. A stepwise adjustments strategy was implemented for learning rate, adjusting every 7 steps with a gamma multiplier of 0.1. Based on prior experience, the maximum training epoch was set to 100, and the batch size was determined to be 16 [28]. After completing 100 training iterations, the training process was halted, and the model was saved.

$$Cros{s}_{entropy}=-\sum _{i=1}^n(P_i\times log{Q}_i)$$

(1)

where i represents the pixel number, n denotes the total pixel number, P represents the measured value, and Q indicates the predicted value.

2.6. Model Evaluation

To evaluate the performance of the Color-Unet model, comparisons were made with FCN, SegNet, DeepLabv3, and the original Unet. Commonly used evaluation metrics in semantic segmentation include mean intersection over union (MIOU) and pixel accuracy (PA), among others. Given that the image segmentation task was a binary classification problem, MIOU, PA, and the F1-score were employed to evaluate the model performance in this study. Additionally, the number of parameters and time cost were considered to evaluate the model’s size and segmentation speed. These metrics were calculated as follows:

$$PA={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}\times 100\%$$

(2)

$$Precision={\displaystyle \frac{TP}{TP+FP}}\times 100\%$$

(3)

$$Recall={\displaystyle \frac{TP}{TP+FN}}\times 100\%$$

(4)

$$F1\text{-}score={\displaystyle \frac{2\times Precision\times Recall}{Precision+Recall}}\times 100\%$$

(5)

$$MIOU={\displaystyle \frac{1}{2}}\times \left({\displaystyle \frac{TP}{TP+FP+FN}} + {\displaystyle \frac{TN}{TN+FN+FP}}\right)\times 100\%$$

(6)

where TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively.

2.7. Segmentation Performance Evaluation

To evaluate the segmentation performance of the Color-Unet model, a comparison was made with the SDTS-MF model [10]. Based on the segmentation results, color grading was performed. Confusion matrices, widely used tool for performance analysis [30], were employed to visually present the model’s segmentation results.

2.8. Seed Cotton Color Measurement

2.8.1. Machine Vision Measurement

Color evaluation was conducted under standard laboratory conditions in accordance with ASTM D5867. Initially, all samples underwent a standardized treatment. Subsequently, each sample was evenly placed into the sample box, and the programmable logic controller operated the electric actuator to compress the sample, ensuring consistent density and surface flatness. Following this, the color measurement of seed cotton was taken. After the measurement, the sample was thoroughly mixed and recompacted for subsequent measurement. Each sample underwent this detection process five times, and the mean values of Rd and +b were calculated.

Seed cotton color measurement comprises four main steps, as illustrated in Figure 8: image acquisition, color calibration, image segmentation, and color index calculation and correction.

• Image acquisition

Images for each seed cotton sample were acquired using the camera of the measurement system. After each image acquisition, the image was numbered and cropped to 1200 × 900 pixels for further processing.

2.

Color calibration

Due to variations in the color temperature of the light source and performance of camera, the seed cotton images captured by the measurement system inevitably exhibited some degree of color deviation and difference. Therefore, it is essential to perform white balance and color difference correction on the images. In this study, a white balance algorithm based on white–black plate and a color correction algorithm based on multiple linear regression were employed, adapted from the method proposed by Li, H. et al. [10].

3.

Image segmentation

Impurity and shadow pixels in images affect the accuracy of seed cotton color index calculations. Therefore, segmenting the seed cotton image is essential to eliminate the influence of the impurities and shadow pixels. In this study, the trained Color-Unet model was utilized to segment the seed cotton images.

4.

Color index calculation and correction

After image segmentation, images containing only pure seed cotton fiber pixels were obtained, which can then be used for seed cotton color index calculations. The specific steps for color index calculation and correction are as follows:

First, the image underwent a nonlinear transformation based on Equations (7)–(10) [31].

$$gamma(x)=\left\{\begin{array}{c}1.055x^{1/2.4}-0.055,(x\ge 0.0031308)\\ 12.92x,(0\le x\le 0.0031308)\end{array}\right.$$

(7)

$$R=gamma({\displaystyle \frac{r}{255}})$$

(8)

$$G=gamma({\displaystyle \frac{g}{255}})$$

(9)

$$B=gamma({\displaystyle \frac{b}{255}})$$

(10)

Second, the spectral tri-stimulus values were calculated according to Equation (11), utilizing a 2° field of view and D65 reference white light.

$$\left(\begin{array}{c}X\\ Y\\ Z\end{array}\right)=\left(\begin{array}{l}0.4124564 0.3575761 0.1804375\\ 0.2126729 0.7151522 0.0721750\\ 0.0193339 0.1191920 0.9503041\end{array}\right)\left(\begin{array}{l}R\\ G\\ B\end{array}\right)$$

(11)

Third, Rd and +b were determined using Equations (12)–(14), and the means of Rd and +b were determined to serve as the final color indices of the seed cotton.

$$Rd=100\overline{Y}/Y_n$$

(12)

$$f_y=0.510\times (21+20\overline{Y}/Y_n)/(1+20\overline{Y}/Y_n)$$

(13)

$$+b=70f_y\times (\overline{Y}/Y_n-0.8470\overline{Z}/Z_n)$$

(14)

where the $\overline{X}$, $\overline{Y}$, and $\overline{Z}$ were the mean of X, Y, and Z, respectively, and the values of X_n, Y_n, and Z_n were 95.047, 100, and 108.883, respectively [10,15].

Finally, the final Rd and +b values were corrected using a correction algorithm based on the BP Neural Network proposed by Li et al. [10].

2.8.2. Spectrophotometer Measurement

After the machine vision measurement, impurities were manually removed from all seed cotton samples. The samples were then treated according to ASTM D5867 standards. Subsequently, the Rd and +b of each sample were measured using HX-410 spectrophotometer under standard laboratory conditions, and these results served as a reference for further analysis.

2.9. Evaluation of Cotton Color Measurement Method

To evaluate the accuracy of the proposed color measurement method, a regression analysis was conducted comparing the outcomes of this method with those obtained from the HX-410 measurement. The evaluation indicators used were R², RMSE, and MRE, which can be calculated using Equations (15)–(17).

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{\sum {(y_i-\overline{y})}^2}}$$

(15)

$$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{n}}}$$

(16)

$$MRE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\displaystyle \frac{\left|y_i-{\widehat{y}}_i\right|}{{\widehat{y}}_i}}\times 100\%$$

(17)

where $y_i$ and $\widehat{y_i}$ represent color index value of the i-th sample from n samples obtained by the proposed method and the HX-410 measurement, respectively.

3. Results

3.1. Training of the Color-Unet Model

Figure 9 illustrates the model training process. The loss curves for training and validation datasets nearly overlap (Figure 9a), and both accuracy trends are consistent (Figure 9b). The loss value initially decreased rapidly, followed by a slower decline, while the accuracy value rose quickly at first before leveling off. After approximately 20 training epochs, both the loss and accuracy values stabilized. Subsequent training iterations showed convergence in the accuracy and loss curves for both datasets, with no signs of overfitting, indicating satisfactory training results. At this stage, the validation dataset achieved an accuracy of 97%, with a loss of 0.5%.

3.2. Segmentation Performance

The Color-Unet achieved a PA of 97.20%, surpassing the PA of the FCN, SegNet, DeepLabv3, and Unet by 3.33%, 3.62%, 3.34%, and 0.52%, respectively (

Table 1. The test results for different models.

Model	PA (%)	F1-Score (%)	MIOU (%)	Time Cost (ms)	Parameters (107)
FCN	93.87	85.28	83.33	121.50	2.01
SegNet	93.58	84.03	82.42	232.20	2.95
DeepLabv3	93.86	85.00	83.10	276.30	3.38
Unet	96.68	90.35	89.66	533.50	3.69
Color-Unet	97.20	93.09	91.81	322.30	1.41

). Additionally, it attained an F1-score of 93.09%, outperforming the F1-score of the FCN, SegNet, DeepLabv3, and Unet models by 7.81%, 9.06%, 8.09%, and 2.74%, respectively. The MIOU for Color-Unet was 91.81%, which exceeded the MIOU of the FCN, SegNet, DeepLabv3, and Unet by 8.48%, 9.39%, 8.71%, and 2.15%, respectively. In terms of model complexity, the Color-Unet has 1.41 × 10⁷ parameters, which is lower than the FCN, SegNet, DeepLabv3, and Unet by 0.7 × 10⁷, 1.54 × 10⁷, 1.97 × 10⁷, and 2.28 × 10⁷, respectively. Furthermore, the Color-Unet segmented each image in an average of 322.3 ms, which is 211.2 ms faster than that of Unet. Thus, the Color-Unet demonstrated both superior performance and reduced processing time compared to other models.

Figure 10 presents raw seed cotton images, manually labeled images, and segmentation outcomes from various models. In Figure 10c, the FCN model produced very rough segmented edges, with numerous small impurities remaining unsegmented. Figure 10d,e show that while SegNet and DeepLabv3 suppressed some noise, they still failed to segment several small impurities. In Figure 10f, the Unet demonstrated effective noise suppression, resulting in smoother segmented edges and fewer unsegmented tiny impurities. Notably, the segmentation results from the Color-Unet closely matched the manually labeled images, further demonstrating that the Color-Unet model exhibited superior segmentation performance compared to the other models.

3.3. Ablation Studies

To evaluate the performance of the CBAM and IIEM in the Color-Unet model, ablation studies were conducted using the seed cotton dataset. The training strategy remained consistent across all models. The segmentation results of different models are summarized in

Table 2. Performance comparison of different improved models.

Model	PA (%)	F1-Score (%)	MIOU (%)	Time Cost (ms)	Parameters (107)
Unet	96.68	90.35	89.66	533.54	3.69
IIEM-Unet	96.87	91.68	90.55	382.17	2.53
CBAM-Unet	96.94	92.02	90.83	326.04	2.33
Color-Unet	97.20	93.09	91.81	322.30	1.41

. Replacing some convolutions in the decoder of the Unet model with IIEM enhanced the segmentation performance, and the PA, F1-score, and MIOU of the IIEM-Unet model increased by 0.19%, 1.33%, and 0.89%, respectively, compared with those of the Unet model. Similarly, replacing some convolutions in the encoder of the Unet model with CBAM enhanced the segmentation performance, too, and the PA, F1-score, and MIOU of the CBAM-Unet model increased by 0.26%, 1.67%, and 1.17%, respectively, compared with those of the Unet model. Notably, when some convolutions were replaced by both the CBAM in the encoder and the IIEM in the decoder, the segmentation performance was significantly improved. The PA, F1-score, and MIOU of the Color-Unet model increased by 0.52%, 2.74%, and 2.15%, respectively, compared with those of the Unet model. Notably, the Color-Unet model outperformed the other models in terms of segmentation speed and the number of parameters, and the Color-Unet model achieved the best segmentation performance. Therefore, the proposed approach effectively integrates attention mechanisms and convolutional decomposition into deep learning frameworks for seed cotton segmentation.

3.4. Comparison with the SDTS-MF

In Figure 11a, both the Color-Unet and SDTS-MF models showed good performance in segmenting impurities and shadows in light-spotted cotton fiber images. However, when segmenting light yellow and yellow stained cotton fibers, the SDTS-MF model exhibited over-segmentation, as shown in Figure 11b,c. The areas of over-segmentation are highlighted with red ovals in these figures. In contrast, the Color-Unet did not experience such over-segmentation, emphasizing its advantage in effectively addressing this issue. This result underscored the superior segmentation accuracy of Color-Unet, particularly in handling color variations without introducing errors.

The comparison of segmentation results between the Color-Unet and SDTS-MF models (Figure 12) reveals notable differences in performance across various seed cotton samples. For 5 light spotted seed cotton samples, both models successfully segmented 3 samples, indicating comparable effectiveness in handling impurities and shadows. However, the performance diverges when evaluating the segmentation of stained samples. Among 27 light yellow-stained seed cotton samples, the SDTS-MF model identified 21, while the Color-Unet achieved a higher count of 25. For the 13 yellow-stained seed cotton samples, the SDTS-MF model recognized only 10, whereas the Color-Unet correctly identified all 13. These results highlighted the superior segmentation capabilities of the Color-Unet, which achieved a remarkable correct recognition rate of 91.11%, significantly surpassing the SDTS-MF model’s rate of 75.56% by 15.55%. This underscores the effectiveness and robustness of the Color-Unet model in accurately segmenting light yellow- and yellow-stained seed cotton images, demonstrating its potential for improving seed cotton quality assessment.

3.5. Color Measurement Performance of Seed Cotton Color Measurement System

The regression analysis comparing the color measurement results from the proposed system and the HX-410 spectrophotometer reveals strong correlations for both the Rd and +b metrics. For Rd, the coefficients were R² = 0.832, RMSE = 1.56%, and MRE = 1.84% (Figure 13a). In contrast, the +b values exhibited even stronger correlations with R² = 0.883, RMSE = 0.150, and MRE = 2.61% (Figure 13b). These result findings indicated that the proposed method for measuring seed cotton color demonstrates a robust relationship with the established HX-410 measurements. The strong R² values signify that a significant portion of the variance in the measurements can be explained by the proposed system, reinforcing the validity and reliability of this new approach in accurately assessing seed cotton color.

4. Discussion

In this study, a machine vision-based (MVT) seed cotton color measurement system was developed to evaluate seed cotton color indexes. MVT effectively combines the subjective strengths of manual inspection with the objectivity and precision of machine -based measurements [14,32]. Compared to the traditional manual detection, the proposed method demonstrates significant advantages in objectivity and efficiency [33]. By eliminating the effects of impurities and shading, it achieves higher identification accuracy than conventional colorimeters and HVI detection methods [34,35,36]. Compared to previous image measurement algorithms [10,15], the proposed algorithms demonstrated superior performance in segmenting shadows and impurities, effectively addressing the over-segmentation issues associated with the SDTS-MF algorithm (as illustrated in Figure 11b,c). This method incorporates calibration algorithms and color index computation techniques to accurately derive color indexes of from RGB images. Notably, the R² for the Rd and +b obtained using the proposed method were 0.023 and 0.004 higher, respectively, than those achieved by the SDTS-MF algorithm. These results highlighted the enhanced accuracy of this approach in measuring seed cotton color, confirming its reliability and precision for providing consistent color assessments.

Impurities and shadows are significant obstacles to achieving accurate seed cotton color detection [6]. Thomasson et al. [12] demonstrated that impurity removal could potentially enhance cotton color indexes. Xu et al. [34] found that reducing shadows and impurities could enhance cotton color grading. To address these challenges, the Color-Unet was developed, extending Unet’s capabilities in this study. Color-Unet not only improved execution time over Unet but also surpassed other models, including FCN, SegNet, DeepLabv3, and Unet (Table 1). These improvements stem from reducing the number of convolutions and the incorporation of CBA and IE modules. Given the complexity of seed cotton images, segmentation was treated as a binary classification task to separate cotton from non-cotton. The Color-Unet model was optimized to be lightweight and fast by reducing the number of convolutions in both the encoder and decoder stages. The CBAM and IIEM modules further boosted performance, especially for segmenting shadows and tiny impurity, as illustrated in Figure 10f. Attention mechanisms like CBAM direct the model to relevant regions, minimizing irrelevant background influence to improve accuracy. CBAM is commonly used in challenging contexts for target extraction [11,27]. The IIEM module decomposes large convolutions into multiple smaller ones, reducing parameters and enhancing segmentation accuracy by increasing depth and width [29]. Ablation studies verified the effectiveness of the CBAM and IIEM in enhancing performance (Table 2). Consequently, the Color-Unet emerged as a lightweight, high-performing model well-suited for rapid, convenient, and accurate seed cotton color measurement.

This study has several limitations. The proposed method for measuring seed cotton color was evaluated under standard conditions, with all samples being maintained at a constant temperature and humidity for 24 h to achieve moisture equilibrium, as per ASTM D5867. This requirement may limit the applicability of the method in non-standard environments. Previous studies demonstrated that moisture content influences cotton color detection [37,38]. Therefore, future work will focus on investigating the effects of various moisture regains on seed cotton color measurements, aiming to enhance the robustness and universality of this method.

5. Conclusions

This study introduced a color detection system for seed cotton based on MVT and proposed a Color-Unet model to detect seed cotton color using RGB images, achieving rapid and accurate detection of seed cotton color. The main findings are as follows:

(1) The Color-Unet model rapidly and accurately segments shadows and impurities in seed cotton images. The PA, F1-score, and MIOU of the Color-Unet model were 97.20%, 93.09%, and 91.81%, which were 0.52%, 2.74%, and 2.2% higher than the Unet model, respectively. The results of the ablation studies confirmed the validity of the CBAM and IIEM. The segmentation speed of the Color-Unet model could be increased, and the number of parameters could be decreased significantly. The average segmentation time per image and the number of parameters of the Color-Unet model were 322.3 ms and 1.41 × 10⁷, which were 211.2 ms and 2.28 × 10⁷ less than that of the Unet model, respectively.

(2) The Color-Unet model successfully addressed the over-segmentation issue present in the SDTS-MF algorithm, which is based on traditional threshold segmentation. The correct recognition rate of Color-Unet model reached 91.11%, surpassing the SDTS-MF model by 15.55%.

(3) The detection system enabled quick and precise measurement of seed cotton color indexes. The R² of the regression between the Rd of this system measurement and that of the HX-410 spectrophotometer was 0.832, the RMSE was 1.56%, and the MRE was 1.84%. The R² of the regression of the +b between the two was 0.883, the RMSE was 0.15, and the MRE was 2.61%. These findings provide essential technical support for seed cotton color detection.