Real-time scene parsing through object detection running on an embedded device is very challenging, due to limited memory and computing power of embedded devices. To deal with these challenges, we redesign a lightweight network without notably reducing detection accuracy. Based on the Darknet-53, we use depth separable convolutions and pointwise group convolutions to reduce the parameter size of the network. A feature extraction backbone network with a parameter size of only 16 percent of darknet-53 is constructed. Meanwhile, in order to compensate for the degradation of accuracy, we have added a Multi-Scale Feature Pyramid Network based on a simple U-shaped structure to improve the performance of multi-scale object detection, which called it Mini-YOLOv3. It has smaller model size and fewer trainable parameters and floating point operations (FLOPs) in comparison of YOLOv3. We evaluate Mini-YOLOv3 on MS-COCO benchmark dataset; The parameter size of Mini-YOLOv3 is only 23% of YOLOv3 and achieves comparable detection accuracy as YOLOv3 but only requires 1/2 detect time, Specifically, Mini-YOLOv3 achieves mAP-50 of 52.1 at speed of 67 fps.

At present, there are two main problems with fruit-and vegetable-picking robots. One is that complex scenes (with backlighting, direct sunlight, overlapping fruit and branches, blocking leaves, etc.) obviously interfere with the detection of fruits and vegetables; the other is that an embedded platform needs a lighter detection method due to performance constraints. To address these problems, a fast tomato detection method based on improved YOLOv3-tiny is proposed. First, we improve the precision of the model by improving the backbone network; second, we use image enhancement to improve the detection ability of the algorithm in complex scenes. Finally, we design several groups of comparative experiments to prove the rationality and feasibility of this method. The experimental results show that the f1-score of the tomato recognition model proposed in this paper is 91.92%, which is 12% higher than that of YOLOv3-tiny; the detection speed on a CPU can reach 25 frames/s, and the inferential speed is 40.35 ms, equivalent to that of YOLOv3-tiny. Through comparative experiments, we can see that the comprehensive performance of our method is better than that of other state-of-the-art methods.

The automatic extraction of lake water is one of the research hotspots in the field of remote sensing image processing. Due to the small interclass variance between lakes and other ground objects, and the complex texture characteristics of lake boundaries, existing methods often have problems such as over-segmentation and inaccurate boundary segmentation when segmenting lake water bodies. To alleviate these problems, this article designs an end-to-end semantic segmentation network [noise-canceling transformer network (NT-Net)] for the automatic extraction of lake water bodies from remote sensing images. Aiming at the problem of over-segmentation caused by nonlake objects, an interference attenuation module is designed in the network. This module can model the key features that are distinguishable and suitable for segmenting lake water by analyzing the difference in feature representation between lakes and other ground objects, thus suppressing the feature representation of nonlake objects. To more accurately segment the lake boundary, a multilevel transformer module is designed. This module can capture the context association of boundary information and enhance the feature representation of boundary information by using the self-attention mechanism. The comparative experimental results show that, compared with the current mainstream semantic segmentation networks, the method in this article has advantages in extracting lake water bodies comprehensively and coherently.

Aiming at the situation that complex natural scene text is difficult to recognize a scene text recognition method based on Encoder-Decoder framework is proposed. The method converts the natural text recognition into a sequence mark by combining the CTC-Attention mechanism under the Encoder-Decoder framework; In order to overcome the problem of character segmentation, using the correlation between image and text sequence. First of all, CNN is used to generate an ordered feature sequence from the entire word image. Then, the generated feature sequence is feature-coded using Bi-LSTM (bidirectional long-term memory) network. Finally, an integrated module of CTC (connection time classification) and Attention mechanism is designed to decode and output the text sequence. Experiments show that compared with the comparison method, the recognition accuracy of the method is improved obviously.