As shown in Figure 2 , DeepLabv3+ is a semantic segmentation model developed by Google that employs an encoder–decoder architecture to integrate multi-scale contextual features and fine spatial details. It uses the Xception network (Chollet, 2017) as its feature extraction backbone and integrates an ASPP module, which combines a 1×1 convolution with three parallel 3×3 dilated convolutions at different dilation rates to capture multi-scale contextual information. The decoder fuses high-level semantic features with low-level edge information, enabling accurate pixel-wise segmentation and enhancing boundary delineation in small objects and complex scenes.

Although DeepLabv3+ exhibits strong performance in semantic segmentation, certain limitations remain that hinder its broader applicability. The encoder reduces the spatial resolution of feature maps and fails to fully exploit high-resolution information, leading to discontinuities in predictions and a loss of fine-grained details. While the ASPP module enhances boundary feature extraction, it inadequately captures local structural information, particularly for radially distributed objects such as weeds and maize leaves, resulting in fragmented segmentation and semantic gaps. Moreover, the Xception backbone introduces a substantial computational burden, increasing hardware demands and reducing training efficiency. To address these limitations, this study proposes a lightweight improved DSC-DeepLabv3+ model that optimizes the encoder–decoder architecture and enhances both computational efficiency and segmentation accuracy. An overview of the proposed model is illustrated in Figure 3 . In the encoder, the original Xception backbone is replaced with MobileNetV2 to achieve efficient feature extraction. The ASPP module is enhanced with depthwise separable convolutions to reduce parameters and improve training efficiency. To further enrich global and local contextual information, a strip pooling module is integrated, resulting in a modified S-ASPP structure with six parallel branches. Additionally, the CBAM is incorporated to enhance segmentation accuracy while maintaining low computational complexity. On the decoder side, the C-CFF module is employed to fuse 1/8 and 1/16 scale feature maps extracted from the backbone. CBAM is introduced again to suppress redundant noise and alleviate edge blurring. Shallow features are fused and upsampled via bilinear interpolation to restore spatial details. The fused deep and multi-scale features are subsequently processed through a 3×3 convolution and fourfold upsampling to restore the original image resolution.

The original DeepLabv3+ employed the Xception network as its feature extraction backbone. Although effective on large-scale datasets, its performance degrades in practical applications characterized by limited annotated data. Its large parameter count and complex architecture result in slow inference and extended training time, hindering deployment on mobile and embedded platforms. To address these limitations, this study proposes a lightweight model by adopting MobileNetV2 (Sandler et al., 2018) as the backbone. MobileNetV2 was introduced by Google in 2018, featuring a streamlined architecture optimized for mobile and embedded devices. It was designed to offer an efficient solution for various visual recognition tasks, including object classification and image segmentation. The architectural details of the backbone network are presented in Table 1 . The core innovation of MobileNetV2 lies in replacing standard convolutions with depthwise separable convolutions. This approach significantly reduces computational cost and model size by decomposing standard convolutions into two operations: depthwise convolution and pointwise convolution. In this structure, the depthwise operation applies a 3×3 convolution to each input channel independently, while the pointwise operation performs a 1×1 convolution to aggregate information across channels and produce the final feature map. This design improves training efficiency and reduces the computational cost of 3×3 convolutions by up to 90%, while maintaining high segmentation accuracy. However, despite these improvements, feature extraction remains suboptimal due to kernel sparsity, with a large proportion of convolutional parameters contributing minimally. To address these issues, MobileNetV2 introduces the linear bottleneck and inverted residual structure, as illustrated in Figure 4 . The linear bottleneck reduces parameter count and computational overhead by eliminating nonlinear activations that may distort low-dimensional feature representations. The inverted residual structure, incorporating designs with two different strides, facilitates network deepening, prevents the vanishing gradient problem, and decreases the number of parameters. Collectively, these innovations allow MobileNetV2 to achieve high performance while maintaining low computational complexity.

This study enhances training efficiency by introducing depthwise separable dilated convolutions, which integrate the benefits of both dilated and depthwise separable operations, as illustrated in Figure 5 . In the depthwise stage, each channel of the feature map is convolved independently with a dilated kernel to extract spatial correlations while preserving local structural information. Subsequently, a 1×1 pointwise convolution aggregates the features across channels to generate the final output. This method enlarges the receptive field, reduces parameter count and computational complexity, and accelerates inference. Additionally, the global average pooling in the traditional ASPP (He et al., 2015) employs a fixed-size square window, which presents certain limitations. Such a design struggles to capture directional scale correlations when processing irregularly shaped objects or complex environments. The square pooling window may introduce redundant dependencies and incorporate noise from unrelated regions, leading to the loss of critical fine-grained information. To overcome these limitations, Strip Pooling (Hou et al., 2020) is integrated into the ASPP module, resulting in a novel six-branch S-ASPP structure. This modification reduces model complexity, enhances inference speed, and enables the capture of diverse multi-scale contextual features. Strip pooling performs directional pooling along horizontal and vertical dimensions to simultaneously capture global and local contextual information while suppressing background noise. It utilizes one-dimensional convolutions along each direction within a residual framework to enhance directional sensitivity. The resulting feature maps are fused and element-wise multiplied with the original features to refine spatial representations. Unlike traditional square pooling, strip pooling independently processes vertical and horizontal spatial dimensions by performing weighted averaging across rows and columns. The structure of the strip pooling module is illustrated in Figure 6 .

For the given input image, the two vectors of the input picture are computed using Equations 1, 2:

For an input X ∈ ℝ C × H × W , the total amount of channels is denoted by C , i , j denotes the row and column respectively, H , W represents the size, X is passed via both vertical and horizontal paths for pooling. The horizontal and vertical outputs are y v ∈ ℝ C × H × W . Following their combination, the Equation 3 is used to figure out the final output:

A refined feature map is obtained after completing the convolution operations and applying the sigmoid activation function, and then it is merged with the input feature map for the production of z . The specific process is shown in Equation 4:

The purpose of this paper is to keep a lightweight model while obtaining outstanding segmentation accuracy. We evaluate the model’s performance using widely utilized semantic segmentation metrics such as mIoU. The calculation of the mIoU is defined by Equation 10:

where n is the number of classes, TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

Parameters reflect the size of the model. FLOPs (Floating Point Operations) are its computational complexity. FPS measures processing speed, representing the time taken to process a picture.

The semantic segmentation model was implemented in the PyTorch framework under the following software environment: Python 3.8.19, PyTorch 2.4.0, CUDA 11.8, and Windows 11. The experiments were conducted on a workstation equipped with an AMD 7745HX CPU and an NVIDIA GeForce RTX 4060 GPU.

During training, input images were resized to 512×512 pixels. Stochastic Gradient Descent (SGD) optimizer was adopted as the optimizer, with an initial learning rate of 0.007, a minimum learning rate set to 0.01 of the maximum, and a weight decay of 0.0001. The training process lasted for 300 epochs, with the first 100 epochs conducted under a frozen backbone using a batch size of 8, and the remaining 200 epochs under an unfrozen backbone with a batch size of 4. Model validation and checkpoint saving were performed every 20 epochs. In the dataset, maize occupies the majority of the pixel area, resulting in significant foreground-background class imbalance. To mitigate the adverse effects of this imbalance on segmentation performance, the cross-entropy loss function was employed, as defined in Equation 11:

where N is the overall count of samples, M symbolizes how many classes, y i c is the true value of the i-th sample belonging to class c, and p i c is the model’s projected likelihood that the i-th sample falls into class c.

Thanks to the pre-trained backbone adopted through transfer learning, the training loss converged rapidly to a low value. Figure 9 depicts the mIoU and loss curves of the proposed model. After 300 epochs, the mean Intersection over Union (mIoU) reached 85.47%. As shown in the loss curve, the validation and training losses decreased to approximately 0.1 and 0.2, respectively, with minimal fluctuations, indicating stable convergence. Beyond this point, further training yielded marginal improvements in loss, suggesting that the model had reached optimal convergence.

To evaluate the performance of the proposed method in maize field weed recognition, we compared it with both classical and lightweight models, including SegNet, BiSeNet, and ICNet. All models were trained under identical experimental conditions and preprocessing procedures. As shown by the loss curves in Figure 10 , the proposed model demonstrates a faster convergence rate and achieves lower final loss values within 300 training epochs for both the training and validation sets. Compared to other models, it also shows a steeper initial loss descent and faster overall convergence, effectively reducing the required training time.

This study compares the accuracy and computational complexity of various models, as summarized in Table 2 . The proposed model reduced FLOPs to 15.326G, which is approximately 150G less than the original DeepLabv3+, while incurring only a 0.71% decrease in mean Intersection over Union (mIoU). Its parameter count was reduced to just 5% of the original, and inference speed increased by 24.49 FPS, making it well-suited for mobile deployment. Compared to SegNet and ICNet, the proposed model achieved mIoU improvements of 7.67% and 5.92%, respectively. It also demonstrated notable gains in inference speed (33.44 and 12.59 FPS), while reducing parameters by 26.55M and 23.61M, and FLOPs by 111G and 12.974G. Compared to BiSeNet, it achieved slightly higher inference speed and a 7.1% increase in mIoU, striking a balance among model compactness, computational efficiency, and segmentation performance.

Figure 11 presents the segmentation results of various models on the maize weed dataset. The proposed model demonstrates superior performance in boundary delineation and pixel-level classification, particularly in scenarios involving overlapping maize and weeds with incomplete or ambiguous shape features. In contrast, models such as PSPNet, FCN, BiSeNet, and ICNet exhibit misclassification of adjacent pixels, primarily due to insufficient global contextual modeling. For instance, the pyramid pooling module in PSPNet compromises spatial detail, while FCN’s limited receptive field overly emphasizes local features, leading to errors in blurred boundary regions. Although BiSeNet and ICNet achieve faster inference through multi-branch architectures, their aggressive downsampling and coarse feature fusion reduce semantic consistency, particularly affecting sensitivity to small objects. U-Net and DeepLabv3+ also suffer from imprecise edge segmentation. In U-Net, skip connections introduce noise from shallow layers, which, when fused with deep features, contribute to edge blurring. Standard upsampling operations further smooth the boundaries, undermining the recovery of fine-grained details. Although the ASPP module in DeepLabv3+ expands the receptive field, its reliance on standard convolutions and global pooling reduces responsiveness to high-frequency edge features. Additionally, the lack of targeted enhancement mechanisms during low-resolution feature fusion in the decoder leads to boundary deviations from the actual object contours. In contrast, the proposed model incorporates strip pooling to capture long-range contextual dependencies and mitigate pixel misclassification in overlapping regions. Furthermore, a feature fusion module enhances the recovery of essential spatial details for precise boundary localization. This architecture effectively balances segmentation accuracy and model efficiency, making it highly suitable for real-time maize weed recognition in complex field environments.

The PASCAL VOC 2012 dataset is widely used in computer vision and includes 21 semantic classes, such as car, person, cat, and dog. It serves as a standard benchmark for evaluating semantic segmentation models. We evaluated the performance of the proposed DSC-DeepLabv3+ model on this dataset to assess its generalization capability. As illustrated in Figure 12 , the model achieves competitive segmentation results. As detailed in Table 5 , our model achieves higher mIoU compared to several existing methods, including U-Net, FCN, PSPNet, and BiSeNet. Although the mIoU is slightly lower than that of the original DeepLabv3+ with an Xception backbone, our model demonstrates significant advantages in terms of parameter count, computational cost, and inference speed. Specifically, compared with the MobileNetV2-based DeepLabv3+, our model improves mIoU by 1.72%, reduces the number of parameters by 50%, lowers computational cost by 73%, and increases inference speed by 11.88 FPS. Furthermore, compared to BiSeNet, our model achieves 3.69% higher mIoU, reduces parameters by 2.89M, and lowers FLOPs by 34.9G. Compared to ICNet, it achieves 3.42% higher mIoU, with a reduction of 23.61M in parameters and 12.9G in FLOPs. Although the inference speed is 6.05 FPS lower than that of BiSeNet, the proposed model achieves a favorable balance between segmentation accuracy and model efficiency. Overall, the experimental results demonstrate the strong generalization capability of the proposed model.