PaperlayerView exact revision on arXiv

Source-linked AI summary

UltraLight Luma: A Novel Edge-Deployable Perception Network for Crop-Row Segmentation in Agricultural Robotics

Dhanushka Balasingham, Shamod Ginigaddarage, Hanojhan Rajahrajasingh, Nuwan Kodagoda, Rajitha de Silva

arXiv:2610.11764v1cs.RO

TL;DR

Low-cost agricultural robots need crop-row perception within tight computation, memory, power, and inference-time limits. UltraLight Luma combines compact RGB-image segmentation with Triangle Scan Method navigation-line extraction; it reports a 74.81% average crop-row detection score using 13.1k parameters.

  • Problem

    Reliable crop-row detection for agricultural-robot navigation must fit low-cost platforms constrained by computation, battery capacity, hardware cost, memory, power, and inference time.

    Source: §I · p. 1

  • Method

    UltraLight Luma segments crop-row regions in RGB images, and the Triangle Scan Method processes predicted masks to estimate crop-row navigation lines.

    Source: §I · p. 1, C. TSM-Based Crop Row Analysis · p. 4

  • Results

    Despite a maximum IoU of 20.68%, UltraLight Luma achieved a 74.81% average crop-row detection score and a 91.0% parameter-efficiency score with 13.1k parameters.

    Source: §V · p. 6

  • Takeaways & Limitations

    UltraLight Luma offers a practical balance between segmentation performance, navigation-line extraction, and edge-deployable efficiency for low-cost agricultural robots.

    Source: §V · p. 6

    The paper reports a modest maximum IoU of 20.68%.

    Source: §V · p. 6

Abstract

from arXiv · show

Reliable crop-row perception is essential for autonomous agricultural robots, but low-cost deployment is constrained by computation, memory, power, and inference time. This paper addresses crop-row detection and navigation-line extraction using UltraLight Luma, a compact encoder-decoder segmentation network with a progressive learning strategy and only 13.1k trainable parameters. The predicted crop-row masks are processed using a post-processing algorithm to estimate row alignment for downstream visual servoing. UltraLight Luma improves parameter efficiency over U-Net, YOLOv8, and YOLOv26 while maintaining reliable crop-row detection performance. The model required only 25.23 mJ per inference, demonstrating its suitability for resource-constrained agricultural robots.

I. INTRODUCTION

Crop-row perception must work reliably while meeting the computation, battery, memory, power, and inference-time limits of low-cost agricultural robots. UltraLight Luma targets this gap with a compact segmentation network and a post-processing module for navigation-line estimation.

I. INTRODUCTION

  • The work targets reliable crop-row detection despite shadows, weeds, missing plants, uneven soil, and perspective distortion, on embedded platforms limited in memory, power, and inference time.The paper frames detection as image segmentation from forward-facing RGB images.Source: §I · p. 1
  • The proposed UltraLight Luma encoder–decoder uses a progressive learning strategy and only 13.1k trainable parameters for crop-row segmentation on low-cost agricultural robots.The paper also applies the Triangle Scan Method to segmented masks to estimate the crop-row navigation line.Source: §I · p. 1

II. RELATED WORK

  • Traditional image-processing methods are efficient but sensitive to illumination and field variability, while YOLO-based approaches can incur centroid-extraction and line-fitting errors under canopy overlap and weed interference.The paper presents semantic segmentation as pixel-wise crop-row localization for row-geometry estimation.Source: §II · p. 1
  • Earlier lightweight crop-row approaches remained above 0.1 million parameters, whereas UltraLight Luma uses 13.1k parameters.The related-work discussion identifies LinkNet, ENet, YOLO-based models, and MobileNet-based backbones among approaches whose parameter counts remained relatively high for extremely constrained deployment.Source: §II · pp. 1–2

III. METHODOLOGY

The methodology passes forward-facing RGB images through segmentation and processes the resulting masks to estimate navigation lines. It evaluates the pipeline using a labeled dataset of field images prepared for training, validation, and testing.

III. METHODOLOGY

  • RGB images pass to UltraLight Luma for crop-row segmentation; the resulting mask is processed by the Triangle Scan Method to estimate angular error for a visual-servoing controller.The complete pipeline generates velocity commands, and the work focuses on crop-row detection and navigation-line extraction for a straddling robot.Source: §III · p. 2, Figure 1 · p. 2
  • Crop-row detection is formulated as binary semantic segmentation, with crop-row pixels labeled 1 and background pixels labeled 0.The predicted mask serves as a perception output for downstream row following and navigation.Source: §III · p. 2

A. Dataset Preparation

  • The dataset contains 2,000 images across two crops, 11 field variations, and 50 data classes, split into 1,250 training, 250 validation, and 500 test images.Images and binary masks are resized to 64×64, with spatial and photometric augmentation used while preserving image–mask correspondence.Source: A. Dataset Preparation · p. 2

B. UltraLight Luma Architecture

UltraLight Luma is a compact encoder–decoder segmentation network designed for edge deployment. Its depthwise-separable blocks reduce parameter and activation-memory demands, and its architecture is organized into four stages.

B. UltraLight Luma Architecture

  • The network replaces standard convolutions with depthwise-separable blocks to reduce parameters and activation memory for edge deployment.Each block applies a 3 × 3 depthwise convolution followed by a 1 × 1 pointwise convolution, with Batch Normalization and ReLU after each step.Source: B. UltraLight Luma Architecture · p. 2
  • The depthwise convolution applies one filter per input channel, while the pointwise convolution mixes information across channels to produce the output feature map.Batch Normalization and ReLU follow each convolution in the composite block.Source: B. UltraLight Luma Architecture · p. 2
  • The network processes inputs through four sequential stages: encoder, downsample, decoder, and output head.Table I summarizes the UltraLight Luma layers.Source: B. UltraLight Luma Architecture · p. 2, Table I · p. 2

1) Encoder:

The encoder extracts and fuses features at multiple scales, then retains an encoder feature map for the decoder. Successive downsampling stages produce a compact bottleneck while preserving skip connections.

1) Encoder:

  • The encoder reduces a 64 × 64 × 3 input to 32 × 32 × 16, fuses features from two spatial scales, and stores the Enc-1 output as skip connection s1.One fusion branch preserves resolution; the other downsamples and upsamples before concatenation and 1×1 pointwise fusion.Source: 1) Encoder · p. 2, Figure 2 · p. 3

2) Downsample:

  • Two successive strided depthwise-separable blocks reduce the representation from 32 × 32 through 16 × 16 to an 8 × 8 bottleneck, retaining a skip connection at 16 × 16.The bottleneck feature map is passed to the decoder; the subscript s = 2 denotes strided convolution for spatial downsampling.Source: 2) Downsample · pp. 2–3

3) Decoder:

The decoder restores spatial resolution by progressively combining upsampled features with encoder skip connections. Its final feature map is converted into per-pixel class probabilities.

3) Decoder:

  • Three bilinear upsampling steps restore spatial resolution, while skip connections and depthwise-separable blocks recover crop-row boundary detail.At each stage, features are concatenated with the corresponding skip connection and fused by a 1×1 pointwise convolution before a DW-Sep block.Source: 3) Decoder · p. 3
  • The decoder uses the previous stage’s feature map and corresponding skip connection; Up(·) denotes bilinear upsampling by a factor of 2.The final d0 feature map is 64×64×16 and is produced by a third upsampling step without a skip connection.Source: 3) Decoder · p. 3
  • A 1×1 convolution maps the final 64×64×16 feature map to two class logits, then softmax produces per-pixel class probabilities.The classes are background and crop-row.Source: 4) Output · p. 3

4) Output:

The output head produces per-pixel probabilities for background and crop-row classes. Training uses a combined focal and Dice loss, with Adam, cosine learning-rate decay, and progressive resolution training.

4) Output:

  • The 1×1 convolution’s class logits are converted to per-pixel probabilities for the two classes, background and crop-row.The output passage identifies zc as the logit for class c and C = 2 as the total number of classes.Source: 4) Output · p. 3
  • The predicted probability at each pixel represents its likelihood of belonging to the background or crop-row class.The output head uses two classes.Source: 4) Output · p. 3

C. Training Objective

  • The combined focal and Dice loss addresses class imbalance by emphasizing hard foreground pixels and encouraging overlap with ground-truth masks.The total loss is defined in Eq. 8.Source: C. Training Objective · p. 3
  • Progressive resolution training moves from 64×64 inputs to higher resolutions and reduces total training time by 28%.Training uses Adam with an initial learning rate of 10^-3 and cosine decay; checkpoints are selected by validation crop-row IoU.Source: C. Training Objective · p. 3

D. Inference and Post-Processing

At inference, UltraLight Luma converts an RGB image into a thresholded crop-row mask, which is skeletonized for row-geometry analysis. Evaluation includes predictions across training stages and quantitative and qualitative assessments under selected field conditions.

D. Inference and Post-Processing

  • The 64×64 crop-row probability channel is upsampled to the original image resolution and thresholded at τ = 0.5 to produce a binary mask.The RGB input is resized to 64×64 and the model produces a 64×64×2 probability tensor.Source: D. Inference and Post-Processing · p. 3
  • Fig. 3 presents predicted crop-row regions in white across training stages from 50 to 300 epochs.The panels correspond to 50, 100, 150, 200, 250, and 300 epochs.Source: Figure 3 · p. 4
  • Skeletonization converts the binary mask to single-pixel-wide row centerlines, followed by dilation for visualization while preserving geometric location.The resulting mask is used for TSM-based analysis.Source: D. Inference and Post-Processing · p. 4

IV. EXPERIMENTAL EVALUATION

  • Evaluation uses IoU across training epochs and qualitative predictions under selected field conditions, with experiments on a Ryzen 7 CPU and NVIDIA RTX 4050 GPU.The RTX 4050 has 4 GB of memory; the evaluation also includes TSM-based analysis and comparisons with YOLO and U-Net where appropriate.Source: §IV · p. 4

A. Training Progress and Segmentation Accuracy

Crop-row IoU increased with training duration, while TSM-based comparisons showed Luma’s higher log-normalized parameter-efficiency score than U-Net and YOLO variants. Qualitative predictions covered land and greenhouse crops, including shadowed conditions.

A. Training Progress and Segmentation Accuracy

  • IoU increased from 15.63% after 50 epochs to 20.68% after 300 epochs; mean training time across six runs was 5.22 seconds per epoch.Total training time for 300 epochs was 26.08 minutes.Source: A. Training Progress and Segmentation… · p. 4

B. Crop Row Prediction Under Diverse Field Conditions

C. TSM-Based Crop Row Analysis

  • TSM fixes the centerline top at the perspective vanishing point, scans bottom-edge candidates, and selects the line with most white pixels to measure angle and lateral offset.Source: C. TSM-Based Crop Row Analysis · p. 4
  • 91.0% average η for Luma, versus 57.6% for U-Net, 57.4% for YOLOv8, and 39.1% for YOLOv26.Luma’s average raw ε was lower than U-Net’s; η provides a log-normalized parameter-efficiency comparison.Source: Figure 5 · p. 5, C. TSM-Based Crop Row Analysis · p. 5
  • 74.81% average ε for Luma, lower than U-Net’s raw ε; log-normalized η supports comparison across models with differing parameter counts.Source: C. TSM-Based Crop Row Analysis · p. 5

D. Computational Efficiency Analysis

UltraLight Luma used less resource than YOLOv8n across four inference-efficiency metrics. Its parameter count was approximately 243 times smaller.

D. Computational Efficiency Analysis

V. CONCLUSION

UltraLight Luma combines crop-row segmentation with Triangle Scan Method-based navigation-line extraction for low-cost agricultural robots. The conclusion reports its segmentation, detection, and parameter-efficiency results.

V. CONCLUSION

  • 20.68% maximum IoU accompanied a 74.81% average crop-row detection score, using 13.1k trainable parameters and a 91.0% parameter-efficiency score.The paper characterizes the maximum IoU as modest and describes the overall balance as practical for resource-constrained agricultural robots.Source: §V · p. 6
  • The predicted crop-row mask is combined with Triangle Scan Method-based navigation-line extraction to support downstream visual servoing.Source: §V · p. 6
  • The approach is presented as balancing segmentation performance, navigation-line extraction, and edge-deployable efficiency for resource-constrained agricultural robots.Source: §V · p. 6

Related papers

Loading 2610.11764v1…