AI-Mandel: CNN-Based Almond Detection

• AI-Mandel is a specialized deep learning pipeline for almond vs. shell classification using a purpose-built 20-layer CNN architecture achieving over 99% accuracy.
• It integrates rigorous preprocessing, augmentation, and regularization techniques to enhance binary discrimination in grayscale images.
• The system provides a core module for automated almond grading and paves the way for future extensions such as multi-class grading and real-time inspection.

“AI-Mandel” is an Editor's term for the automatic image annotation and almond-detection workflow centered on AlmondNet-20, the purpose-built 20-layer convolutional neural network introduced in “Automatic Image Annotation (AIA) of AlmondNet-20 Method for Almond Detection by Improved CNN-based Model” (Ilani et al., 2024). In the source paper, the system is defined by a binary visual discrimination task: distinguishing whole almonds from empty shells in grayscale images. The reported objective is to enhance the grading process for almonds and their shells through Deep Convolutional Neural Networks, with AlmondNet-20 trained over 1,000 epochs and evaluated as achieving overall accuracy above 99% in the reported setting (Ilani et al., 2024).

1. Problem setting and scope

The source work situates AlmondNet-20 within the grading of agricultural products, particularly almonds and shells, in response to a “burgeoning global demand for premium agricultural products” and the “competitive nut market” (Ilani et al., 2024). Its immediate task is not general-purpose agrifood inspection, but a narrowly specified binary classification problem: distinguishing whole almonds from empty shells.

Within that scope, the paper presents the system as an Automatic Image Annotation (AIA) workflow for almond detection. The reported implementation operates on grayscale imagery and is designed to improve reliability and accuracy in classification. The paper further states that the resulting classification system offers “tangible benefits to both industry experts and non-specialists alike,” ensures “globally reliable almond classification,” and creates opportunities for “product patents,” while contributing to economic value through improved agrifood quality control (Ilani et al., 2024).

A plausible implication is that AI-Mandel, in this formulation, should be understood less as a general theory of almond phenotyping than as a specialized visual classification stack for a tightly constrained industrial inspection problem.

2. AlmondNet-20 architecture

AlmondNet-20 is explicitly described as a purpose-built 20-layer CNN designed from the ground up for almond-versus-shell discrimination in grayscale images (Ilani et al., 2024). The input is a $210 \times 320 \times 1$ image. The full layer sequence reported in the paper is as follows.

Layer	Operation	Parameters
Input	Grayscale image	$210 \times 320 \times 1$
1	Conv2D	64 filters, kernel $3 \times 3$, stride $1 \times 1$, padding = “same”, activation = ReLU
2	MaxPooling2D	pool_size = $2 \times 2$, stride = $2 \times 2$
3	Conv2D	128 filters, kernel $3 \times 3$, stride $1 \times 1$, padding = “same”, activation = ReLU
4	MaxPooling2D	$2 \times 2$, stride = $2 \times 2$
5	SpatialDropout2D	rate = 0.1
6	Conv2D	512 filters, kernel $210 \times 320 \times 1$0, stride $210 \times 320 \times 1$1, padding = “same”, activation = ReLU
7	MaxPooling2D	$210 \times 320 \times 1$2, stride = $210 \times 320 \times 1$3
8	Conv2D	256 filters, kernel $210 \times 320 \times 1$4, stride $210 \times 320 \times 1$5, padding = “same”, activation = ReLU
9	MaxPooling2D	$210 \times 320 \times 1$6, stride = $210 \times 320 \times 1$7
10	Conv2D	256 filters, kernel $210 \times 320 \times 1$8, stride $210 \times 320 \times 1$9, padding = “same”, activation = ReLU
11	MaxPooling2D	pool_size = $3 \times 3$0, stride = $3 \times 3$1
12	Conv2D	128 filters, kernel $3 \times 3$2, stride $3 \times 3$3, padding = “same”, activation = ReLU
13	MaxPooling2D	$3 \times 3$4, stride = $3 \times 3$5
14	Conv2D	128 filters, kernel $3 \times 3$6, stride $3 \times 3$7, padding = “same”, activation = ReLU
15	MaxPooling2D	$3 \times 3$8, stride = $3 \times 3$9
16	Dropout	rate = 0.5
17	Flatten	—
18	Dense	64 units, activation = ReLU
19	BatchNormalization	—
20	Dense	2 units, activation = Softmax

The paper states that every convolutional layer uses a ReLU non-linearity to introduce sparsity, and that each MaxPooling stage halves or thirds the spatial dimensions to build hierarchical feature maps (Ilani et al., 2024). SpatialDropout2D and Dropout are interspersed to guard against over-fitting, while BatchNormalization is placed at the penultimate stage to stabilize the dense feature space before the final Softmax classifier.

The architecture combines progressively abstracted spatial features with dense classification layers. This suggests that the model is intended to capture subtle geometric and textural distinctions rather than relying on a shallow feature hierarchy.

3. Mathematical formulation

The paper provides three central expressions: convolution, categorical cross-entropy loss, and accuracy (Ilani et al., 2024). The convolution operation is reported as

$1 \times 1$0

The categorical cross-entropy loss is given as

$1 \times 1$1

The reported accuracy metric is

$1 \times 1$2

In this formulation, the Softmax output layer produces a two-class posterior over almond and shell, while training minimizes categorical cross-entropy. The use of ReLU activations, dropout-based regularization, and BatchNormalization situates AlmondNet-20 within a standard supervised CNN classification regime, but the paper emphasizes that its effectiveness derives from the combination of depth, augmentation, and preprocessing rather than from any single component.

4. Data handling, preprocessing, and training protocol

To prevent memorization of the training set, the authors applied on-the-fly augmentation via Keras’s ImageDataGenerator (Ilani et al., 2024). The reported augmentations are:

• random rotations up to $1 \times 1$3,
• horizontal and vertical flips,
• random zooming in/out by up to $1 \times 1$4,
• brightness shifts $1 \times 1$5.

The paper states that this synthetically expanded dataset diversity and enforced invariance to orientation, scale, and lighting. In addition, the reported preprocessing pipeline includes grayscale conversion, Gaussian blur, adaptive thresholding, and Canny edges.

The training protocol is specified in detail. The model was trained for 1,000 epochs with Adam, using learning rate $1 \times 1$6, $1 \times 1$7, and $1 \times 1$8, and a batch size of 32 (Ilani et al., 2024). Regularization consisted of $1 \times 1$9 weight decay $2 \times 2$0, SpatialDropout2D with rate 0.1, Dropout with rate 0.5, and BatchNormalization. Class balancing used scikit-learn class weights “computed to counter the 1 : 1 almond:shell ratio.” Training was accelerated on two GPUs in Kaggle, reducing the 1,000-epoch runtime by approximately 40 minutes compared with CPU only.

Taken together, these choices define AI-Mandel as a full training and inference pipeline rather than a bare network topology. The paper attributes robustness not only to the CNN depth but also to the interaction among augmentation, regularization, and preprocessing.

5. Reported performance and robustness

On a held-out test set of 106 images, AlmondNet-20 is reported to have yielded the following metrics (Ilani et al., 2024):

• Precision$2 \times 2$1, Recall$2 \times 2$2, $2 \times 2$3
• Precision$2 \times 2$4, Recall$2 \times 2$5, $2 \times 2$6
• Overall accuracy = 0.991 (99.1%)
• Final training loss = 0.0567
• Validation loss = 0.0605

The confusion matrix reported in the paper is:

True \ Predicted	Shell	Almond
Shell	44	0
Almond	0	12

The paper states that the model never confused almonds with shells in the test set, that no misclassifications were recorded, and that both precision and recall reached unity (Ilani et al., 2024). The authors attribute this result to “balanced augmentation, class weighting, and deep hierarchical features,” together with “careful pre-processing (grayscale conversion, Gaussian blur, adaptive thresholding, Canny edges)” and the depth of AlmondNet-20, which is said to learn “subtle shape, texture, and boundary cues differentiating the edible kernel from its brittle shell.”

A common misconception would be to treat these results as evidence that the system already solves the broader problem of almond quality assessment. The paper itself does not make that claim. It reports strong performance for a binary detection setting and then explicitly identifies several directions required for broader deployment.

6. Extensions and interpretive context

The conclusion proposes five future extensions for the system (Ilani et al., 2024). First, the authors propose generalizing from binary detection to multi-class grading, including “broken, chipped, discolored kernels.” Second, they propose applying the same architecture to different almond varieties in order to capture varietal shape and color differences. Third, they propose incorporating multispectral or near-infrared imaging to detect moisture content or internal defects. Fourth, they propose embedding a lightweight version of AlmondNet-20 into a real-time conveyor-belt inspection system or an orchard-mounted UAV. Fifth, they propose augmenting the base model with attention mechanisms or transformer modules for finer detail discrimination.

These extensions clarify the present status of AI-Mandel. In its reported form, it is a binary CNN-based visual classifier optimized for almond-versus-shell discrimination in grayscale images. The proposed future work suggests an intended trajectory toward broader grading, richer sensing modalities, and real-time deployment. A plausible implication is that the current system functions as a core detection module on which more general agrifood inspection systems could be built, rather than as a complete end-state solution.

Within that interpretation, AlmondNet-20 exemplifies a class of domain-specific CNN architectures in which carefully chosen preprocessing, aggressive augmentation, and explicit regularization are combined with a moderately deep classifier to achieve high task-specific accuracy. The paper presents this combination as a step toward “fully automated, high-volume almond grading systems” with improved speed, reliability, and objective consistency (Ilani et al., 2024).