Rice Plaque Detection and Identification Based on an Improved Convolutional Neural Network

Rice diseases are extremely harmful to rice growth, and achieving the identification rapid classification of disease spots is an essential means promote intelligent production. However, due large variety similar appearance some diseases, existing deep learning methods less effective at detection. Aiming such problems, this paper took spot images five common as research object constructed a data set containing 2500 bacterial blight, sheath flax leaf spot, streak blast, including 500 each disease. An improved lightweight network model was proposed realize accurate types spots. A image designed based on RlpNet (rice plaque net) model, Which underlying network, in addition YOLOv3 target detection order achieve optimization feature extraction link, i.e., upsampling by transposed convolution downsampling dilated convolution. The compared with traditional convolutional neural models, AlexNet, GoogLeNet, VGG-16 ResNet-34 for recognition, results showed that average recall, precision, F1-score overall accuracy were 91.84%, 92.14%, 91.87% respectively, which all algorithms. FSSD, Faster-RCNN, YOLOv4 studies, it could mean precision (mAP) 86.72%, rate (DR) 93.92%, frames per second (FPS) 63.4 false alarm (FAR) only 5.12%. In summary, comprehensive performance better than algorithm, so study provides new method It also had good terms multiple support differentiation has practical application value.